Fractal image compression finds an image-to-image mapping (iterative function system) whose fixed point is the target image. This problem is recognized as Nearest Subspace Search (NSS) problem, because the primitive mapping from Domain image block to Range block has two parameters of intensity magnification and offset. These parameters span 2D subspace. In this paper, we first present that this NSS problem is equivalent to Nearest Neighbor Search (NNS) problem, because all subspaces involve a common non-zero vector. Essentially equivalent interpretation has been shown in D.~Saupe's past research. This paper, however, shows more simple and general interpretation. Based on this interpretation, we compared equivalent NNS settings with different query vectors and clarify the most efficient setting. In addition, based on this interpretation, we generalize the fractal image compression.

Various systems have been proposed to support biological image analysis, with the intent of decreasing false annotations and reducing the heavy burden on biologists. These systems generally comprise a feature extraction method and a classification method. Task-oriented methods for feature extraction leverage characteristic images for each problem, and they are very effective at improving the classification accuracy. However, it is difficult to utilize such feature extraction methods for versatile task in practice, because few biologists specialize in Computer Vision and/or Pattern Recognition to design the task-oriented methods. Thus, in order to improve the usability of these supporting systems, it will be useful to develop a method that can automatically transform the image features of general propose into the effective form toward the task of their interest. In this paper, we propose a semi-supervised feature transformation method, which is formulated as a natural coupling of principal component analysis (PCA) and linear discriminant analysis (LDA) in the framework of graph-embedding. Compared with other feature transformation methods, our method showed favorable classification performance in biological image analysis.

本論文では,Gaussian Process Regression(GPR)に基づく異常予兆検出用のモニタリング法を提案する.異常予兆検出はプラントや,生体,乗り物等の健全性モニタリングのために必要であり,幅広い用途がある.異常予兆検出は,正常事例の蓄積をしながら,蓄積された事例に基づいて異常予兆の検出を行う問題であり,オンライン処理が前提となる.事例に基づく異常予兆の検出は,非線型回帰計算であるGPRによってセンサ値を推定し,推定値と実測値の乖離を調べることで実現できる.しかし,GPRでは蓄積された事例のペアを引数とするカーネル関数から成るグラム行列の逆行列を求めるため,事例数の3乗のオーダの計算が必要になる.したがって,オンラインで蓄積され続ける全事例を用いたGPRの計算は,時間とともに急激に減速するため,実時間性が要求される問題には適用できない.本論文では,まず,この問題を解決するために,入力に応じて回帰計算に用いる事例集合(Active Set)を動的に絞り込むことによって,精度を保ちながら回帰計算の高速化を実現する方法を提案する.次に,瞬時的な異常から長時間の趨勢の異常までをモニタリングするために,様々な時間解像度での異常度を表すSpectro Anomaly Gramを提案する.これにより,瞬時的な異常から大域的な異常までがモニタリングできるようになる.

本論文では,産業用インクジェットプリンター(IIJP:Industrial Ilnk Jet Printer)の印字検査用OCR(Optical Character Reader)に適した高速かつ高精度な文字切出し法を提案する.筆者らは任意の識別器の振舞いを線形回帰木によって模倣し,これを識別器として使用することで,高速なOCRを構築し,このOCRを用いて,トップダウン的に文字認識と切出しを実行する手法CCS(Classification based Character Segmentation)を提案した.しかしながら,CCSでは印刷された文字が水平かつ等間隔に配置されていることを前提としており,個々の文字の大きさや二次元的な位置の変動には対応できない.この問題を解決するために,局所特徴を用いてボトムアップ的に文字領域候補をあらかじめ抽出しておき,これらの各領域に対する文字認識の評価値の和を最大化する組合せ最適化計算を行うMCCS(Modified CCS)を提案する.局所特徴には,本研究で開発したFast-Hessian-Affine regionsを使用する.実験では,まず様々な局所特徴を比較することでFast-Hessian-Affine-regionsが文字切出しに適した特徴であることを示し,次に多様な二次元的変動をもった印字に対してMCCSを適用した結果から,提案手法の有効性を明らかにする.

本論文では,劣化画像を対象とした検索・照合の問題に適した画像間の相違性尺度を提案する.画像はその観測過程で,陰影の変化,遮へい,ぼけやぶれ,解像度低下など劣化の影響をしばしば受ける.これらの劣化に左右されにくい検索・照合を実現するために,これまでは劣化の影響を受けにくい特徴を使用することが研究されてきた.しかし,全ての劣化に対する不変特徴は存在せず,使用する特徴だけで全ての劣化の影響を緩和あるいは除去することはできない.本研究では,特徴ではなく相違性尺度を変更することによって,劣化の影響を受けにくい画像間相違度を提案する.本論文では,「画像を直交展開した際の項の欠落または混入」によって劣化をモデル化する.このモデルのもとでは,欠落の起きていない部分の展開係数が一致すれば,劣化画像は比較した画像に近いといえる.この考えに基づき,まず,任意の直交展開に対して2枚の画像の展開係数の一致数ができるだけ多くなるように劣化画像の輝度値を調整する.次に,一致する要素が多いベクトル対に対してより小さい評価値が与えられる尺度を,上記の輝度調整された画像の展開係数ベクトルの組に適用する.このようにして得られた画像間相違度を用いた実験を通じて,提案する相違度により照合精度が,従来法を用いる場合よりも大幅に向上することを示す.

マルチメディア情報検索に関しては,これまでにも多くの研究がなされてきたが,特に近年では様々な変化に対して不変な画像の局所特徴量の登場を背景として,画像をクエリとして類似した画像を検索する技術が進歩している.本稿では,情報爆発時代における類似画像検索の基盤技術として,特徴ではなく,それらの間の「類似性尺度」をいかに与えればよいかという問題に焦点を当てる.特に劣化画像をキーとした類似画像検索タスクを通じて,よく用いられる正規化相互相関などよりも極めて優れた検索結果をもたらす類似性尺度に関する最近の研究について紹介し議論する.

一般に高精度な識別器ほど低速であることが知られている.そこで本論文では任意の識別器を線形回帰木によって識別速度を高速化する手法,"Classifier Molding"を提案する.Moldingには高精度な識別器と大量の学習データが必要となる.本論文では,高精度であるが低速である混合類似度法(CSM:Compound Similarity Method)を高速化の対象とし,産業用インクジェットプリンタ(IIJP)による印字の文字認識問題に適用した.まず,CSMの学習データは,IIJPの印字特性を模した摂動を基準パターンに与えることで自動生成し,次に,学習済みのCSMの入出力関係を大量に収集した.この入出力関係を学習データとして線形回帰木を構築する.線形計算と二分探索だけで構成される回帰木における演算は非常に高速であるため,構築した線形回帰木はCSMの振舞いを模倣した高速な識別器とみなすことができる.また,高速性を利用することで,識別に基づいた文字切出し手法CCS(Classification based Character Segmentation)を開発した.CCSは,組合せ最適化の枠組みを採用することによって画像中を効率良く探索し,探索後の合計類似度を最大化できる点列を文字位置として抽出する手法である.実験では,CSMの識別精度を低下させることなく1500倍の高速化に成功したことと,画素の連結成分に依存していた従来手法によるセグメンテーションエラーがCCSによって全て回避できることを示し,提案手法の有効性を明らかにした.

This paper presents a robust and real-time object tracking method in image sequences. Existing visual object detection and tracking algorithms usually use apparent features of target object as a model. Those algorithms detects objects by thresholding the similarity or the dissimilarity measure with the model. They sometimes fail to detect objects because of inadequate threshold. In this paper, we propose a method for detection and tracking using positive (object) and negative (non-object) models. This method does not require any threshold, because the object region can be detected just by comparing the similarities or dissimilarities with positive and negative models. If we employ distance as a dissimilarity measure, the problem can be formalized as a nearest neighbor (NN) classification problem. We first show an example of NN classification based color object detection, which detects multiple objects in real-time. Next, we extend the method to tracking. For this extension, we have to measure the distinctiveness of the object, because the detected object by the NN classification sometimes disappears. By finding the high distinctiveness image region around the tracked region in the previous image frame, we can keep tracking even if the object cannot be detected by NN classification. We implemented this method on an active stereo camera system and confirmed that video-rate detection and tracking as well as 3D position measurement can he realized just by using standard PC.

In the case that the sizes of faces are not available, all possible sizes of faces have to be assumed and a face detector has to classify many (often ten or more) sub-image regions everywhere in an image. This makes the face detection slow and the high false positive rate. This paper explores the usage of depth information for accelerating the face detection and reducing the false positive rate at the same time. In detail, we use the depth information to determine the size of the sub-image region that needs to be classified for each pixel. This will reduce the number of sub-image regions that need to be classified from many to one for one position (pixel) in an image. Since most unnecessary classifications are effectively avoided, both the processing time for face detection and the possibility of false positive can be reduced greatly. We also propose a fast algorithm for estimating the depth information that is used to determine the size of sub-image regions to be classified.

本論文では,K-means trackerにて,より安定かつ高速に対象を追跡するために,背景に応じた非ターゲットクラスタ中心の数とそれらの配置を適応的に求める方法について述べる.提案手法では,K-means trackerにより求められたサーチエリア楕円の輪郭上の画素を走査し,ターゲット,非ターゲットクラスタ中心までの距離をもとに,追跡に都合のよい非ターゲットクラスタ中心の個数,配置を決定する.これにより,従来のK-means trackerの非ターゲットクラスタ中心の配置方法における問題である,似た特徴をもつ画素の多重選択問題と,ターゲットに近い特徴をもつ背景画素の選択もれの問題を同時に解決できる.従来のK-means trackerとの比較実験により,本手法は追跡を安定かつ,効率的に処理できることを確認した.

本論文では,独立して動作する2台の能動カメラを用い,高速に運動している物体(150度/秒程度)を追跡し,対象の鮮明な画像を撮影できる高速追従型2眼能動カメラシステムを提案する.本論文では,鮮明な画像を,適切な大きさで撮影された,ぶれやピントずれがない対象の画像と定義する.能動カメラは,FV-PTZ (Fixed Viewpoint Pan-Tilt-Zoom)カメラを使用し,画像内における対象追跡はK-means trackerアルゴリズムを用いる.提案システムでは,追跡対象を注視するために,「人間の両目」のように2台の能動カメラを協調させ,視線を追跡対象上の1点で交わらせるように能動カメラを制御する.これを実現するために,K-means trackerに信頼度の概念を導入し,相対的な信頼度を用いた2台の能動カメラの制御方法を提案する.この信頼度とエピポーラ拘束を利用することにより,それぞれの能動カメラの追従性能を保ったまま,矛盾のないカメラの首振り制御を行い,また,ズーム・フォーカスも同時に制御することにより,より鮮明な対象の画像の撮影し続ける能動追跡システムを実現する.

This paper presents an RK-means clustering algorithm which is developed for reliable data grouping by introducing a new reliability evaluation to the K-means clustering algorithm. The conventional K-means clustering algorithm has two shortfalls: 1) the clustering result will become unreliable if the assumed number of the clusters is incorrect; 2) during the update of a cluster center, all the data points belong to that cluster are used equally without considering how distant they are to the cluster center. In this paper, we introduce a new reliability evaluation to K-means clustering algorithm by considering the triangular relationship among each data point and its two nearest cluster centers. We applied the proposed algorithm to track objects in video sequence and confirmed its effectiveness and advantages.

This paper presents a general constructive algorithm to recover external camera parameters from a set of pairwise partial camera calibrations embedded in the structure named Camera Dependency Graph (CDG) [1] that encompasses both the feasibility and the reliability of each calibration. An edge in CDG and its weight account for the existence and for the quality of the essential matrix between the two views connected by it, respectively. Any triplet of cameras sharing visible points forms a triangle in a CDG, which permits to compute the relative scale between any two of its edges. The algorithm first selects from CDG the set of feasible paths being the shortest ones in terms of reliability that also are connected by a sequence of triangles. The global external parameters of the arrangement of cameras are computed in a process in two steps that aggregates partial calibrations, represented by triangles, along the paths connecting pairs of views taking into account the relative scales between triangles until recovering the parameters between the extremes of each path. Finally, the scales of the whole set of paths are referred to one canonical value corresponding to the edge in the CDG working as the global scale. Initial experimental results on simulated data demonstrate the usefulness and accuracy of such scheme that can be applied either alone or as the initial approximation for other calibration methods.

Viola et al. have introduced a fast object detection scheme based on a boosted cascade of haar-like features. In this paper we introduce a novel ternary feature that enriches the diversity and the flexibility significantly over haar-like features. We also introduce a new genetic algorithm (GA) based method for training effective ternary features through iterations of feature generation and selection. Experimental results showed that the rejection rate can reach at 98.5% with only 16 features at the first layer of the constructed cascade detector This indicates the high performance of our method for generating effective features. We confirmed that the training time can be significantly shortened compared with Violas's method while the performance of the resulted cascade detector is comparable to the previous methods.

This paper brings out a robust pixel-wise object tracking algorithm which is based on the K-means clustering algorithm. In order to achieve the robust object tracking under complex condition (such as wired objects, cluttered background), a new reliability-based K-means clustering algorithm is applied to remove the noise background pixel (which is neigher similar to the target nor the background samples) from the target object. According to the triangular relationship among an unknown pixle and its two nearest cluster centers (target and background), the normal pixel (target or background one) will be assigned with high reliability value and correctly classified, while noise pixels will be given low reliability value and ignored. A radial sampling method is also brought out for improving both the processing speed and the robustness of this algorithm. According to the proposed algorithm, we have set up a real video-rate object tracking system. Through the extensive experiments, the effectiveness and advantages of this reliability-based K-means tracking algorithm are confirmed.

This paper brings out a robust pixel-wise object tracking algorithm which is based on the K-means clustering algorithm. In order to achieve the robust object tracking under complex condition (such as wired objects, cluttered background), a new reliability-based K-means clustering algorithm is applied to remove the noise background pixel (which is neigher similar to the target nor the background samples) from the target object. According to the triangular relationship among an unknown pixle and its two nearest cluster centers (target and background), the normal pixel (target or background one) will be assigned with high reliability value and correctly classified, while noise pixels will be given low reliability value and ignored. A radial sampling method is also brought out for improving both the processing speed and the robustness of this algorithm. According to the proposed algorithm, we have set up a real video-rate object tracking system. Through the extensive experiments, the effectiveness and advantages of this reliability-based K-means tracking algorithm are confirmed.

Viola et al. have introduced a fast object detection scheme based on a boosted cascade of haar-like features. In this paper we introduce a novel ternary feature that enriches the diversity and the flexibility significantly over haar-like features. We also introduce a new genetic algorithm based method for training effective ternary features. Experimental results showed that the rejection rate can reach at 98.5% with only 16 features at the first layer of the cascade detector We confirmed that the training time can be significantly shortened while the performance of the resulted cascade detector is comparable to the previous methods.

The Particle filter estimates a probability, distribution of target object's state by sampled hypotheses and their weights. This method is more expressive than existing method such as Kalman filtering, because the object state is represented as a multi-modal distribution. However the method can't be directly applied to temporally variable appearance object tracking, for example, a firefly or a flickering neon-sign. For solving this problem, we propose a particle filter for a variable appearance object, which estimates a unique state parameter independent of target's position. Our method decomposes the state space into disjoint parameter spaces, i.e., object position and posture space and Appearance parameter space. In the appearance parameter space, the likelihood of each hypothesis is evaluated at the position parameters generated in the other space, and the best explain parameter is determined. Based on this parameter likelihood in the position and posture space is evaluated. By interacting the parameter estimations in different spaces, we can successfully track blinking fire-fly in the darkness.

Hua et al. have proposed a stable and efficient tracking algorithm called "K-means tracker" [2, 3, 5]. This paper describes an adaptive non-target cluster center selection method that replaces the one used in K-means tracker where non-target cluster center are selected at fixed interval. Non-target cluster centers are selected from the ellipse that defines the area for searching the target object in K-means tracker by checking whether they have significant effects for the pixel classification and are dissimilar to any of the already-selected nontarget cluster centers. This ensures that all important non-target cluster centers will be picked up while avoiding selecting redundant non-target clusters. Through comparative experiments of object tracking, we confirmed that both the robustness and the processing speed could be improved with our method.

In this paper, we present a general object tracking method based on a newly proposed pixel-wise clustering algorithm. To track an object in a cluttered environment is a challenging issue because a target object may be in concave shape or have apertures (e.g. a hand or a comb). In those cases, it is difficult to separate the target from the background completely by simply modifying the shape of the search area. Our algorithm solves the problem by 1) describing the target object by a set of pixels; 2) using a K-means based algorithm to detect all target pixels. To realize stable and reliable detection of target pixels, we firstly use a 5D feature vector to describe both the color ("Y, U, V") and the position ("x, y") of each pixel uniformly. This enables the simultaneous adaptation to both the color and geometric features during tracking. Secondly, we use a variable ellipse model to describe the shape of the search area and to model the surrounding background. This guarantees the stable object tracking under various geometric transformations. The robust tracking is realized by classifying the pixels within the search area into "target" and "background" groups with a K-means clustering based algorithm that uses the "positive" and "negative" samples. We also propose a method that can detect the tracking failure and recover from it during tracking by making use of both the "positive" and "negative" samples. This feature makes our method become a more reliable tracking algorithm because it can discover the target once again when the target has become lost. Through the extensive experiments under various environments and conditions, the effectiveness and efficiency of the proposed algorithm is confirmed.

This paper presents a new method for tracking the moving target by controlling a pan-tilt camera. Our method can capture motion-blur-free images of a moving target, because this method automatically synchronizes the target motion with the camera motion. We employed Fixed Viewpoint Pan Tilt Camera (FV-PTZ Camera) as an active camera, which does not produce motion parallax by the camera rotation. For the target tracking on image space, we use K-means Tracker algorithm. As for the pan-tilt control, we employ the PID control scheme. This is because if by assigning the P component as the angular velocity of the object, the I and D components can be corresponded to the angular position and the angular acceleration, respectively. This means the PID control is suitable for controlling the angular speed and position of the pan-tilt unit simultaneously. PID based pan-tilt control is effective for the motion synchronization between the target and camera.

This paper presents a new method for simultaneous object tracking and recognition using object image database. This application requires two searches: search for object appearance stored in the database and that for pose parameters (position, scale, orientation, and so on) of the tracking object in each image frame. For simplifying this problem, we propose a new method, pose parameter embedding (PPE) that transforms the original problem to an appearance search problem. The nearest neighbor (NN) appearance search in this problem has a special property that gradually changing queries are given. For this problem, graph based NN search is suitable, because the preceding search result can be used as the starting point of the next search. Delaunay graph can be used for this search, however, both the graph construction cost and the degree (number of mean edges connected to a vertex) drastically increase in high-dimensional space. Instead, we propose nearest first traversing graph (NFTG) for avoiding these problems. Based on these two techniques, we successfully realized video-rate tracking and recognition.

In this paper, we brought out a noise-insensitive pixel-wise object tracking algorithm whose kernel is a new reliable data grouping algorithm that introduces the reliability evaluation into the existing K-means clustering (called as RK-means clustering). The RK-means clustering concentrates on two problems of the existing K-mean clustering algorithm: 1) the unreliable clustering result when the noise data exists; 2) the bad/wrong clustering result caused by the incorrectly assumed number of clusters. The first problem is solved by evaluating the reliability of classifying an unknown data vector according to the triangular relationship among it and its two nearest cluster centers. Noise data will be ignored by being assigned low reliability. The second problem is solved by introducing a new group merging method that can delete pairs of "too near" data groups by checking their variance and average reliability, and then combining them together. We developed a video-rate object tracking system (called as RK-means tracker) with the proposed algorithm. The extensive experiments of tracking various objects in cluttered environments confirmed its effectiveness and advantages.

This paper presents a novel method for human head tracking using multiple cameras. Most existing methods estimate 3D target position according to 2D tracking results at different viewpoints. This framework can be easily affected by the inconsistent tracking results on 2D images, which leads 3D tracking failure. For solving this problem, an extension of CONDENSATION Using multiple images has been proposed. The method generates many hypotheses on a target (human head) in 3D space and estimates the likelihood of each hypothesis by integrating viewpoint dependent likelihood values of 2D hypotheses projected onto image planes. In theory, viewpoint dependent likelihood values should be integrated by multiplication, however, it is easily affected by occlusions. Thus we investigate this problem and propose a novel likelihood integration method in this paper and implemented a prototype system consisting of six sets of a PC and a camera. We confirmed the robustness against occlusions.

In this paper, we propose a high-performance object tracking system for obtaining high-quality images of a high-speed moving object at video rate by controlling a pair of active cameras that consists of two cameras with zoom lens mounted on two pan-tilt units. In this paper, "high-quality image" implies that the object image is in focus and not blurred, the size of the object in the image remains unchanged, and the object is located at the image center. To achieve our goal, we use the K-means tracker algorithm for tracking objects in an image sequence captured by the active cameras. We use the results of the K-means tracker to control the angular position and speed of each pan-tilt-zoom unit by employing the PID control scheme. By using two cameras, the binocular stereo vision algorithm can be used to obtain the 3D position and velocity of the object. These results are used in order to adjust the focus and zoom. Moreover, our system allows the two cameras to gaze at a single point in 3D space. However, this system may become unstable when the time response deteriorates by excessively interfering in a mutual control loop or by strict restriction of the camera action. In order to solve these problems, we introduce the concept of reliability into the K-means tracker, and propose a method for controlling the active cameras by using relative reliability. We have developed a prototype system and confirmed through extensive experiments that we can obtain focused and motion-blur-free images of a high-speed moving object at video rate.

This paper describes a sophisticated method to track iris contour and to estimate eye gaze for blinking eyes with a monocular camera. A 3D eye-model that consists of eyeballs, iris contours and eyelids is designed that describes the geometrical properties and the movements of eyes. Both the iris contours and the eyelid contours are tracked by using this eye-model and a particle filter. This algorithm is able to detect "pure" iris contours because it can distinguish iris contours from eyelids contours. The eye gaze is described by the movement parameters of the 3D eye model, which are estimated by the particle filter during tracking. Other distinctive features of this algorithm are: 1) it does not require any special light sources (e.g. an infrared illuminator) and 2) it can operate at video rate. Through extensive experiments on real video sequences we confirmed the robustness and the effectiveness of our method.

This paper presents a novel method for human head tracking using multiple cameras. Most existing methods estimate 3D target position according to 2D tracking results at different viewpoints. This framework can be easily affected by the inconsistent tracking results on 2D images, which leads 3D tracking failure. For solving this problem, an extension of CONDENSATION using multiple images has been proposed. The method generates many hypotheses on a target (human head) in 3D space and estimates the likelihood of each hypothesis by integrating viewpoint dependent likelihood values of 2D hypotheses projected onto image planes. In theory, viewpoint dependent likelihood values should be integrated by multiplication, however, it is easily affected by occlusions. Thus we investigate this problem and propose a novel likelihood integration method in this paper and implemented a prototype system consisting of six sets of a PC and a camera. We confirmed the robustness against occlusions.

This paper presents object detection and tracking algorithm which can adapt to object color shift. In this algorithm, we train and build multi target models using color under different illumination conditions. Each model called as Color Distinctiveness look up Tables or CDT The color distinctiveness is the value integrating 1) similarity with target colors and 2) dissimilarity with non-target colors, which represents how distinctively the color can be classified into target pixel. Color distinctiveness can be used for pixel-wise target detection, because it takes 0.5 for colors on decision boundary of nearest neighbor classifier in color space. Also, it can be used for target tracking by continuously finding the most distinctive region. By selecting the most suitable CDT for camera direction, lighting condition, and camera parameters. the system can adapt target and background color change. We implemented this algorithm for a Pan-tilt stereo camera system. Through experiments using this system, we confirmed that this algorithm is robust against color shift caused by illumination change and it can measure the target 3D position at video rate.

Background subtraction is widely used as an effective method for detecting moving objects in a video image. However, background subtraction requires a prerequisite in that image variation cannot be observed, and the range of application is limited. Proposed in this research paper is a method for detecting moving objects by using background subtraction that can be applied to cases in which the image has varied due to varying illumination. This method is based on two object detection methods that are based on different lines of thinking. One method compares the background image and the observed image using invariant features of illumination. The other method estimates the illumination conditions of the observed image and normalizes the brightness before carrying out background subtraction. These two methods are complementary, and highly precise detection results can be obtained by ultimately integrating the detection results of both methods. © 2006 Wiley Periodicals, Inc.

In this paper, we present a clustering-based tracking algorithm for tracking people (e.g. hand, head, eyeball, body). A human body often appears as a concave object or an object with apertures. In this case, many background areas are mixed into the tracking target which are difficult to be removed by modifying the shape of the search area during tracking. This algorithm realizes the robust tracking for such objects by classifying the pixels in the search area into "target" and "non-target" with K-means clustering algorithm that uses both the "positive" and "negative" samples. The contributions of this research are: 1) Using a 5 - D feature vector to describe both the geometric feature "(x,y)" and color feature "(Y,U,V)" of an object (or a pixel) uniformly. This description ensures our method to follow both the position and color changes simultaneously during tracking; 2) Using a variable ellipse model (a) to describe the shape of a non-rigid object (e.g. hand) approximately, (b) to restrict the search area, and (c) to model the surrounding non-target background. This guarantees the stable tracking of objects with various geometric transformations. Through extensive experiments in various environments and conditions, the effectiveness and the efficiency of the proposed algorithm is confirmed.

In this paper, we propose a high performance object tracking system for obtaining high quality images of a high-speed moving object at video rate by controlling a pair of active cameras mounted on two fixed view point pan-tilt-zoom units. In this paper, the 'High quality object image' means that the image of the object is in focus and not blurred, the S/N ratio is high enough, the size of the object in the image is kept unchanged, and the position the object appearing at the image center. To achieve our goal, we use K-means tracker algorithm for tracking object in image sequence which taken by the active cameras. We use the result of the K-means tracker to control the angular position and speed of each pan-tilt-zoom unit by employing PID control scheme. By using two cameras, binocular stereo vision algorithm can be used to obtain 3D position and velocity of the object. These results are used for adjust the focus and the zoom. Moreover, our system let two cameras fix its eyes on one point in 3D space. However, this system may be unstable, when time response loses by interfering in a mutual control loop too much, or by hard restriction of cameras action. In order to solve these problems, we introduced a concept of reliability into K-means tracker, and propose a method for controlling active cameras by using relative reliability. We produce the prototype system. Though extensive experiments we confirmed that we can obtain in focus and motion- blur-free images of a high-speed moving object at video rate.

In this paper we present a clustering-based tracking algorithm for non-rigid object. Non-rigid object tracking is a challenging task because the target often appears as a concave shape or an object with apertures. In such cases, many background areas will be mixed into the tracking target, which are difficult to be removed by modifying the shape of the search area. Our algorithm realizes robust tracking for such objects by classifying the pixels in the search area into "target" and "background" with K-means clustering algorithm that uses both the "positive" and "negative" samples. The contributions of this research are: 1) Using a 5D feature vector to describe both the geometric feature "(x, y)" and color feature "(Y, U, V)" of an object (or a pixel) uniformly. This description enables the simultaneous adaptation of both the geometric and color variance during tracking; 2) Using a variable ellipse model (a) to describe the search area; (b) to model the surrounding background. This guarantees the stable tracking of objects with various geometric transformations. Through extensive experiments in various environments and conditions, the effectiveness and the efficiency of the proposed algorithm is confirmed.

We have proposed a K-means clustering based target tracking method, compared with the template matching, which can work robustly <i>when tracking an object with hole through which the background can be seen</i> (e.g., <i>mosquito coil</i>) (<i>hereafter we call this problem as the</i> background interfusion<i>or the</i> interfused background). This paper presents a new method for solving the drawbacks of the previous method, i.e., low speed, instability caused by the change of shape and size. Our new tracking model consists of a single target center, and a variable ellipse model for representing non-target pixels. The contributions of our new method are: 1) The original K-means clustering is replaced by a 2_&infin;-means clustering, and the non-target cluster center is adaptively picked up from the pixels on the ellipse. This modification reduces the number of distance computation and improves the stability of the target detection as well. 2) The ellipse parameters are adaptively adjusted according to the target detection result. This adaptation improves the robustness against the scale and shape changes of the target. Through the various experiments, we confirmed that our new method improves speed and robustness of our original method.

In this paper, we present a clustering-based tracking algorithm for tracking people (e.g. hand, head, eyeball, body, and lips). It is always a challenging task to track people under complex environment, because such target often appears as a concave object or having apertures. In this case, many background areas are mixed into the tracking area which are difficult to be removed by modifying the shape of the search area during tracking. Our method becomes a robust tracking algorithm by applying the following four key ideas simultaneously: 1) Using a 5D feature vector to describe both the geometric feature "(x,y)" and color feature "(Y,U,V)" of each pixel uniformly. This description ensures our method to follow both the position and color changes simultaneously during tracking
2) This algorithm realizes the robust tracking for objects with apertures by classifying the pixels, within the search area, into "target" and "background" with K-means clustering algorithm that uses both the "positive" and "negative" samples. 3) Using a variable ellipse model (a) to describe the shape of a nonrigid object (e.g. hand) approximately, (b) to restrict the search area, and (c) to model the surrounding non-target background. This guarantees the stable tracking of objects with various geometric transformations. 4) With both the "positive" and "negative" samples, our algorithm achieves the automatic self tracking failure detection and recovery. This ability makes our method distinctively more robust than the conventional tracking algorithms. Through extensive experiments in various environments and conditions, the effectiveness and the efficiency of the proposed algorithm is confirmed. © 2006 ACADEMY PUBLISHER.

This paper describes a sophisticated method to estimate visual direction using iris contours. This method requires only one monocular image taken by a camera with unknown focal length. In order to estimate the visual direction, we assume the visual directions of both eyes are parallel and iris boundaries are circles in 3D space. In this case, the two planes where the iris boundaries reside are also parallel. We estimate the normal vector of the two planes from the iris contours extracted from an input image by using an extended "two-circle" algorithm. Unlike most existing gaze estimation algorithms that require information about eye corners and heuristic knowledge about 3D structure of the eye in addition to the iris Contours, our method uses two iris contours only. Another contribution of our method is the ability of estimating the focal length of the camera. It allows one to use a zoom lens to take images and the focal length can be adjusted at any time. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This paper presents a novel learning method for nonlinear mapping between arbitrary dimensional spaces. Unlike artificial neural nets, GMDH, and other methods, our method doesn't require complicated control parameters. Providing a feasible error threshold and training samples, it automatically divides the objective mapping into partially linear mappings. Since decomposed mappings are maintained by a binary tree, the linear mapping corresponding to an input is quickly selected. We call this method Partially Linear Mapping tree (PaLM-tree) . In order to estimate the most reliable linear mappings satisfying the feasible error criterion, we employ split-and-merge strategy for the decomposition. Through the experiments on function estimation, image segmentation, and camera calibration problems, we confirmed the advantages of PaLM-tree.

This paper describes a novel method for digitizing the 3D shape of an object in real-time, which can be used for capturing live sequence of the 3D shape of moving or deformable objects such as faces. Two DMD (digital micro mirror) devices are used as high speed switches for modulating and demodulating light rays. One DMD is used to generate rays of light pulses, which are projected onto the object to be measured. Another DMD is used to demodulate the light reflected from the object illuminated by the light pulses into intensity image that describes the disparity. A prototype range finder implementing the proposed method has been built. The experimental results showed that the proposed method works and video sequences of disparity images can be captured in real time. © 2005 IEEE.

We have proposed a K-means clustering based target tracking method, compared with the template matching, which can work robustly <i>when tracking an object with hole through which the background can be seen</i> (e.g., <i>mosquito coil</i>) (<i>hereafter we call this problem as the</i> background interfusion<i>or the</i> interfused background). This paper presents a new method for solving the drawbacks of the previous method, i.e., low speed, instability caused by the change of shape and size. Our new tracking model consists of a single target center, and a variable ellipse model for representing non-target pixels. The contributions of our new method are: 1) The original K-means clustering is replaced by a 2_&infin;-means clustering, and the non-target cluster center is adaptively picked up from the pixels on the ellipse. This modification reduces the number of distance computation and improves the stability of the target detection as well. 2) The ellipse parameters are adaptively adjusted according to the target detection result. This adaptation improves the robustness against the scale and shape changes of the target. Through the various experiments, we confirmed that our new method improves speed and robustness of our original method.

In this paper, we propose a new feedback-error-learning controller enhanced by the PaLM tree that is an easy-to-use function approximator developed by our research group. We investigate the ability of our feedback-error-learning controller by applying it to controlling an active camera head to pursuit a moving target with high accuracy and high response. The PaLM-tree learns the inverse model in the feedback-error-learning scheme correctly. Although our active camera head has unknown mechanical friction and our closed-loop control system has a relatively large latency, our active camera head can successfully pursuit eye movement by our feedback-error-learning controller based on the PaLM-tree. Through experiments, we confirmed that our method could achieve high-performance control over the tuned feedback control.

Three-dimensional (3-D) video is a real 3-D movie recording the object's full 3-D shape, motion, and precise surface texture. This paper first proposes a parallel pipeline processing method for reconstructing a dynamic 3-D object shape from multiview video images, by which a temporal series of full 3-D voxel representations of the object behavior can be obtained in real time. To realize the real-time processing, we first introduce a plane-based volume intersection algorithm: first represent an observable 3-D,space by a group of parallel plane slices, then back-project observed multiview object silhouettes onto each slice, and finally apply two-dimensional silhouette intersection on each. slice. Then, we propose a method to parallelize this algorithm using a PC cluster, where we employ five-stage pipeline processing in each PC as well as slice-by-slice parallel silhouette intersection. Several results of the quantitative performance evaluation are given to demonstrate the effectiveness of the proposed methods. In the latter half of the paper, we present an algorithm of generating video texture on the reconstructed dynamic 3-D object surface. We first describe a naive view-independent rendering method and show its problems. Then, we improve the method by introducing image-based rendering techniques. Experimental results demonstrate the effectiveness of the improved method in generating high fidelity object images from arbitrary viewpoints.

This paper presents a perspective imaging model that is able to eliminate image deformations caused by non-linear lens distortion and to detect the location of the optical center respectively. By using a calibration pattern it can detect the optical center by tracking the optical rays generated by 3D points that have the same representation on the image plane. Calibration pattern is also used in order to generate normalized virtual screens on which captured images are back projected, thus eliminating the deformations. Experimental results for single viewpoint omni-directional cameras demonstrate the effectiveness of our method. © 2004 IEEE.

In this paper, we describe a novel camera calibration method to estimate the extrinsic parameters and the focal length of a camera by using only one single image of two coplanar circles with arbitrary radius.
We consider that a method of simple operation to estimate the extrinsic parameters and the focal length of a camera is very important because in many vision based applications, the position, the pose and the zooming factor of a camera is adjusted frequently.
An easy to use and convenient camera calibration method should have two characteristics: 1) the calibration object can be produced or prepared easily, and 2) the operation of a calibration job is simple and easy. Our new method satisfies this requirement, while most existing camera calibration methods do not because they need a specially designed calibration object, and require multi-view images. Because drawing beautiful circles with arbitrary radius is so easy that one can even draw it on the ground with only a rope and a stick, the calibration object used by our method can be prepared very easily. On the other hand, our method need only one image, and it allows that the centers of the circle and/or part of the circles to be occluded.
Another useful feature of our method is that it can estimate the focal length as well as the extrinsic parameters of a camera simultaneously. This is because zoom lenses are used so widely, and the zooming factor is adjusted as frequently as the camera setting, the estimation of the focal length is almost a must whenever the camera setting is changed. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This paper describes a conic-based algorithm for estimating visual line from a single monocular image. By assuming that the visual lines of the both eyes are parallel and the iris boundaries are circles, we propose a "two-circle" algorithm that can estimate the normal vector of the supporting plane of the iris boundaries, from which the visual line is calculated. Our new method does not use either the eye corners, or some heuristic knowledge about the structure of the eye. Another advantage of our algorithm is that a camera with an unknown focal length can be used without assuming the orthographical projection. This is a very useful feature because it allows one to use a zoom lens and to change the zooming factor whenever he or she likes. It also gives one more freedom of the camera setting because keeping the camera far from the eyes is not necessary in our method. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This paper describes a novel method to estimate visual line from a single monocular image. By assuming that the visual lines of the both eyes are parallel and the iris boundaries are circles, we propose a "two-circle" algorithm that can estimate the normal vector of the supporting plane of the iris boundaries, from which the direction of the visual line can be calculated. Most existing gaze estimation algorithms require eye corners and some heuristic knowledge about the structure of the eye in addition to the iris contours. In contrast to the exiting methods, our one does not use either of those additional information. Another advantage of our algorithm is that a camera with an unknown focal length can be used without assuming the orthographical projection. This is a very useful feature because it allows one to use a zoom lens and to change the zooming factor whenever he or she likes. It also gives one more freedom of the camera setting because keeping the camera far from the eyes is not necessary in our method, The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

Voronoi condensing reduces training patterns of nearest neighbor classifiers without changing the classification boundaries. This method plays important roles not only in the nearest neighbor classifiers but also in the other classifiers such as the support vector machines, because the resulting prototype patterns involve support vectors in many cases. However, previous algorithms for Voronoi condensing were computationally inefficient in general Pattern Recognition tasks. This is because they use proximity graphs for entire training patters, which require computational time exponentially for the dimension of pattern space. For solving this problem, we proposed an efficient algorithm for Voronoi condensing named direct condensing that does not require the entire proximity graphs of training patterns. We confirmed that direct condensing efficiently calculates Voronoi condensed prototypes in high dimension (from 2 to 20 dimensions).

This paper describes a novel method to estimate visual direction from a single monocular image with "two-circle" algorithm. We assume that the visual direction of both eyes is parallel and iris boundaries are circles in 3-D space. Our "two-circle" algorithm can estimate the normal vector of the supporting plane of two iris boundaries, from which the direction of the visual direction can be calculated. Most existing gaze estimation algorithms require eye corners and some heuristic knowledge about the structure of the eye as well as the iris contours. In contrast to the exiting methods, ours does not use that additional information. Another advantage of our algorithm is that it does not require the focal length, therefore, it is capable of estimating the visual direction from an image taken by an active camera. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This study proposes a method for recognition of the behavior and number of multiple objects without separation of the objects from images. Most conventional techniques of behavior recognition have used bottom-up processing, in which features were first extracted from images, and then the extracted features were subjected to time-series analysis. However, separation of objects from images at the feature extraction stage resulted in unstable processing. This study aims at stable recognition of multiobject behavior. For this purpose, a mechanism of selective attention is proposed. With this mechanism, particular image regions (focusing regions) are allotted to all states of the NFA (nondeterministic finite automaton) that performs sequence analysis, and feature extraction (event detection) is performed inside such regions. This approach makes it possible to detect events irrespective of noise (that is, changes that may occur in the image beyond the focusing regions), while nondeterministic state transition means that all possible event sequences are analyzed
hence, the behavior of multiple objects can be recognized without separation of the objects from the images. Object-specific color tokens are assigned to NFA active state sets, and then are transferred along with the state transitions, which is referred to as the object discrimination mechanism. Introduction of this mechanism allows simultaneous multiobject behavior recognition and detection of the number of objects. In addition, the proposed system has been extended to treat multiview images, and its effectiveness has been proven experimentally. © 2001 Scripta Technica.

Recognizing multiple object behaviors from nonsegmented image sequences is a difficult problem because most of the motion recognition methods proposed so far share the limitation of the single-object assumption. Based on existing methods, the problem can be solved only by bottom-up image sequence segmentation followed by sequence classification. This straightforward approach totally depends on bottom-up segmentation which is easily affected by occlusions and outliers. This paper presents a completely novel approach for this task without using bottom-up segmentation. Our approach is based on assumption generation and verification, i.e., feasible assumptions about the present behaviors consistent with the input image and behavior models are dynamically generated and verified by finding their supporting evidence in input images. This can be realized by an architecture called the selective attention model, which consists of a state-dependent event detector and an event sequence analyzer. The former detects image variation (event) in a limited image region (focusing region), which is not affected by occlusions and outliers. The latter analyzes sequences of detected events and activates all feasible states representing assumptions about multiobject behaviors. In this architecture, event detection can be regarded as a verification process of generated assumptions because each focusing region is determined by the corresponding assumption. This architecture is sound since all feasible assumptions are generated. However, these redundant assumptions imply ambiguity of the recognition result. Hence, we further extend the system by introducing 1) colored-token propagation to discriminate different objects in state space and 2) integration of multiviewpoint image sequences to disambiguate the single-view recognition results. Extensive experiments of human behavior recognition in real world environments demonstrate the soundness and robustness of our architecture.

In a Cooperative Distributed Vision system, a group of communicating Active Vision Agents (AVA, in short, i.e. real time image processor with an active video camera and high speed network interface) cooperate to fulfill a meaningful task such as moving abject tracking and dynamic scene visualization. A key issue to design and implement an AVA rests in the dynamic integration of Visual Perception, Camera Action, and Network Communication. This paper proposes a novel dynamic system architecture named Dynamic Memory Architecture, where perception, action, and communication modules share what we call the Dynamic Memory. It maintains not only temporal histories of state variables such as pan-tilt angles of the camera and the target object location but also their predicted values in the future. Perception, action, and communication modules are implemented as parallel processes which dynamically read from and write into the memory according to their own individual dynamics. The dynamic memory supports such asynchronous dynamic interactions (i.e. data exchanges between the modules) without wasting time for synchronization. This no-wait asynchronous module interaction capability greatly facilitates the implementation of real time reactive systems such as moving object tracking. Moreover, the dynamic memory supports the virtual synchronization between multiple AVAs, which facilitates the cooperative object tracking by communicating AVAs. A prototype system for real time moving object tracking demonstrated the effectiveness of the proposed idea.

Spatial Reasoning, reasoning about spatial information (i.e. shape end spatial relations), is a crucial function of image understanding and computer vision systems. This paper proposes a novel spatial reasoning scheme for image understanding and demonstrates its utility and effectiveness in two different systems: region segmentation and aerial image understanding systems. The scheme is designed based on a so-called Multi-Agent/Cooperative Distributed Problem Solving Paradigm, where a group of intelligent agents cooperate with each other to fulfill a complicated task. The first part of the paper describes a cooperative distributed region segmentation system, where each region in an image is regarded as an agent. Starting from seed regions given at the initial stage, region agents deform their shapes dynamically so that the image is partitioned into mutually disjoint regions. The deformation of each individual region agent is realized by the snake algorithm(14) and neighboring region agents cooperate with each other to find common region boundaries between them. In the latter part of the paper, we first give a brief description of the cooperative spatial reasoning method used in our aerial image understanding system SIGMA. In SIGMA, each recognized object such as a house and a road is regarded as an agent. Each agent generates hypotheses about its neighboring objects to establish spatial relations and to detect missing objects. Then, we compare its reasoning method with that used in the region segmentation system. We conclude the paper by showing further utilities of the Multi-Agent/Cooperative Distributed Problem Solving Paradigm for image understanding.

The image partitioning problem is to find the set of those regions satisfying both their attribute constraints and relational constrains. Attribute constraint means the statistical distribution of the intensity, color and texture of the region, and also the clarity of the region contour edge. Relational constraint is that all regions form disjoint union of the image space. To solve this problem by taking account of both attributes of regions and relations among them, the multi-agent processing is suitable. In this paper, a multi-agent image segmentation method is proposed, where a group of agents representing regions cooperate with each other to partition an image into a set of disjoint regions. The most distinguishing characteristics of the method are : 1. It employs spatial relations between regions as well as their attributes to partition an image. 2. A flexible spatial reasoning is realized by the cooperation among multiple agents. Each agent actively generates hypotheses about its goal state (i.e. expected region boundary). The hypotheses generated by the top-down region boundary estimation and the attributes of a region obtained by the bottom-up analysis are integrated through the energy minimization process. Then, each agent modifies its state (i.e. region deformation)so as to minimize the energy function. The hypothesis generation and region deformation are repeated referring to those hypotheses and states of the other agents ; the inconsistency between hypotheses is examined and less reliable ones are modified to resolve the inconsistency. Some experimental results demonstrate the effectiveness of the proposed method.

The gamma-omega Hollgh transform we proposed earlier (a) does not lead to bias in the number of votes accumulated in the cell even when the parameter space is sampled in uniform cells and voting takes place over all pixels, and (b) the voting locus becomes a piecewise linear line composed of two segments so drawing and analyzing the curve is simple. These are significant advantages. In a conventional gamma-omega Hough transformation algorithm, however, votes from the pixel set included on one digital line spread over multiple cells in the parameter space and the number of pixels forming digital lines is not always taken as the correct number of votes. The reason is the:conventional algorithm does not detect all of the target digital lines in the image space. In this research, for the gamma-omega Hough parameter space, we determine the cell configuration having a one-to-one correspondence with all of the digital lines in the image space and demonstrate an appropriate voting method for this cell configuration. By applying the ''high-precision gamma-omega Hough transformation algorithm'' used in the cell configuration and voting method proposed in this paper, digital lines having any orientation and position can be stably and precisely detected.

我々が以前提案したγ-ωハフ変換は,(a)パラメータ空間を均一なセルで標本化し全画素からの投票を行っても,セルに蓄積される投票度数に偏りが生じない.(b)投票軌跡が2本の線分からなる折れ線となり,軌跡の描画や解析が容易に行える.という優れた特長をもっている.しかし,従来のγ-ωハフ変換アルゴリズムでは,1本のディジタル直線に含まれる画素集合からの投票がパラメータ空間中の複数のセルに分散され,ディジタル直線を構成する画素の数が投票度数として正しくとらえられない場合がある.これは,従来のアルゴリズムが,画像空間中に存在するすべてのディジタル直線を検出対象としていないことに起因する.本研究では,γ-ωパラメータ空間において,画像空間中に存在するすべてのディジタル直線と1対1に対応したセル配置を求め,そのセル配置に対する妥当な投票方法を明らかにする.本論文で提案するセル配置と投票方法を用いた「高精度γ-ωハフ変換アルゴリズム」を用いることにより,任意の方向,位置をもつディジタル直線が安定かつ高精度に検出できる.

The Hough transform is an effective method for detecting figure elements in images corrupted by noise. Voting on the feature points in the image space is performed in the parameter space, after which the points in the parameter space with many votes are focused on in order to detect the figure elements. In the Hough transform, the parameter space is partitioned into elements called cells where the votes are accumulated. However, when the subject is a digital image, if the parameter space is not sampled appropriately, a bias arises in the number of votes accumulated in the cells. In this paper, we present a sampling method for the parameter space that does not produce distortion in the ρ–θ parameter space used in line detection, and we construct a γ–ω space where no bias in the number of votes arises even when sampling is uniform. The γ–ω parameter space corresponds to the ρ–θ parameter space and has the features of no bias in the number of votes and voting loci, which are segmented lines. In addition, the characteristics of the γ–ω parameter space, where the feature points and voting loci are easily converted from one to the other, are used, and a verification method for the line segments is proposed that does not scan the image space again. By combining the line detection method that uses the γ–ω parameter space and this verification method, stable line detection becomes possible. Copyright © 1993 Wiley Periodicals, Inc., A Wiley Company

近年のオープンアクセスジャーナルの流れに沿って登場し，増加の一途を辿っている粗悪学術論文誌と粗悪学術会議について，その背景，見分けるためのチェックポイント，具体的な弊害，ブラックリストの必要性，研究者倫理との関連，対策，等について概説した．

【目的】間接灸は火傷の可能性もあり温度特性について把握しておく必要がある。しかし先行研究 では1秒未満の詳細な温度変化を検討したものは見受けられない。そこで台座灸、温筒灸、棒灸 の温度特性について検討した。<br> 【方法】4㎜厚のシナベニア板上に置いた1㎜厚のアメゴムシート上に熱電対を施灸部位直下（0.00 ㎜）、直下から外方3.75㎜、7.50㎜、15.00㎜、30.00㎜、45.00㎜の6点に設置し、温度インターフェ イスを介してパーソナルコンピュータに温度データを取り込んだ。棒灸は、熱電対からの高さ20 ㎜と100㎜とした。0.55秒間隔で600秒間計測し、各灸6回測定した。<br> 【結果および考察】平均最高温度は台座灸55.9±5.0℃、温筒灸64.3±3.3℃、棒灸の高さ20㎜は 51.2±4.7℃、 100㎜は30.1℃±3.3℃だった。最高温度までの時間は、 台座灸160.7秒、 温筒灸154.5秒、 棒灸の高さ20㎜は126.7秒、100㎜は182.5秒だった。台座灸の温度曲線は漸増的に温度上昇後、 頂点付近は弧を描き、なだらかに温度下降をした。温筒灸は、台座灸よりも急激に温度上昇し、 頂点付近で少しゆるやかになり頂点から急激に下降した。棒灸の高さ20㎜は、漸増後、頂点はゆ るやかな弧をえがき、非常になだらかに直線的に温度は下降、高さ100㎜は非常になだらかに温度 上昇し、直線的に推移した。台座灸、温筒灸では施灸部外方7.5㎜以遠ではほとんど温度上昇がな かった。台座灸では最高温度付近で急激な温度変化があることがわかった。各種間接灸を使い分 けることにより、異なる刺激を与えることができる可能性が示唆された。<br> 【結語】高頻度の温度計測および施灸周囲の温度測定によりこれまで不明だった台座灸、温筒灸、 棒灸の温度特性をとらえることができた。

【目的】棒灸は治療家のみならず、セルフケアとしても用いられているが、火傷の過誤が報告されるなど、温度特性について把握しておく必要がある。棒灸の燃焼部からの高さ、経時的な温度変化、影響範囲について検討した。
【方法】4mm厚のシナベニア板上に置いた1mm厚のアメゴムシート上に熱電対（ST-50 理科工業）を燃焼部直下（0mm）、直下から外方3.75mm、7.5mm、15mm、30mm、45mmの6点に設置し、温度インターフェイス（Ｅ830 テクノセブン）を介してパソコンに温度データを取り込んだ。棒灸（温灸純艾條 カナケン）は、先端から熱電対までの高さを20、30、40、60、80、100mmと変えて10分間経時的に6回測定した。
【結果および考察】測定時の室温は24.5±2.5℃だった。測定点の平均最高温度は、高さ20mmでは44.1±4.7℃、30mmは38.5±2.6℃、40mmは36.2±4.5℃、60mmは29.4±2.9℃、80mmは26.8±2.3℃、100mmは25.6℃±3.1℃だった。高さの100mmの燃焼部からの水平距離では直下から3.75mmでは25.9±3.3℃、7.5mmは25.9±3.2℃、15mmは25.9±3.1、30mmは26.1±3.0℃、45mmは26.3±3.1℃だった。また、高さ20mmで直下と外方部の温度逆転が、260秒後に外方15mm、290秒後に7.5mm、310秒後に3.75mm、430秒後に30mm、600秒後に45mmとの間に見られた。灰による影響と思われる。
【結語】直下では燃焼部に近いほど最高温は高かったが、高さ100mmでは遠いほど温度が高く、経時的には水平距離と温度の関係の逆転もみられた。棒灸は燃焼部からの高さにより、上昇温度、刺激範囲を可変でき、術者が刺激を調節することができることが確認された。

For Multiple Instance Image Retrieval (MIIR), we have proposed index feature reduction method, which reduces the number of indexes attached to the image entries in the database. This method computes the importance measure representing the stability and the discrimination power of each feature by using the framework of Diverse Density and reduces the features having less importance measures. Through the experiments, this reduction drastically reduces the memory usage and retrieval accuracy, but the acceleration of retrieval speed is limited. This is because the number of nearest neighbor searches performed for each query is equivalent to the number of local features in the query image. This means that query feature reduction is required for the acceleration of image retrieval. This report presents a query feature reduction method for solving this problem.

Gaussian process regression is a nonlinear regression that estimates the expected value of the of the output and its variance for given input. The original Gaussian process regression estimates a scalar value as an expected value of output, which can simply be extended to estimate a vector value. However, the covariance matrix cannot be estimated by simple extension. We have proposed an accelerated Gaussian process regression by introducing dynamic active set consisting of input-output pairs, and the weighted output covariance can be utilized as the covariance of output for multivariate Gaussian process regression. However, the diagonal elements of the estimated matrix can be negative. This is caused by the negative weight of the outputs originated by the inverse of gram matrix. In this report, we propose a method to estimate non-negative weights for the outputs. The estimation is twofold: initially estimate the output vector by simple multivariate Gaussian process regression, then recompute non-negative weights so as to minimize the output error. The resulted weights are utilized to estimate covariance matrix. By using the proposed method in anomaly detection of plant data and electrocardiogram data in the experiment, it was confirmed its effectiveness.

This report presents a cooperation method for multiple pedestrian detection systems, each of which has a single camera with wide view angle, for extending their observation area. Most single-viewpoint human detection algorithms perform matching between pedestrian models and captured images. Since the major task of the target system is people counting, the output of the system is rather simple. We assume that only the pedestrian positions on the image plane are available. For the cooperation of such systems, we propose map-based cooperation scheme, where the detected positions by multiple systems are transformed by nonlinear regression onto a single floor map, and the locations detected by different systems are merged based on the proximity. In our research, we employ linear regression tree for the nonlinear regression due to its short computation time. However, single camera people detection system cannot produce accurate floor position without the height information of each pedestrian. This is because the half upper body of the human model has strong matching weight, and hence, the head position is well estimated on the image plane but the foot position is not. Our method computes the correspondences of the locations of the same pedestrian, and the pedestrian's height is estimated simultaneously based on the discrepancy of their locations on the map.

Content Based Image Retrieval (CBIR) using local features can be classified into two types. One is the method using single vector representation that integrates local features detected in an image into a single Bag of Feature (BoF) vector. The other uses multiple instance representation without integration. We call the latter method Multiple Instance Image Retrieval (MIIR). In MIIR, a method for reducing database indexes has been proposed. This method employs the framework of Diverse Density (DD) to represent the importance measure, which means the stability as well as the discriminative power of the feature (instance). This reduction reduces the memory usage and the retrieval accuracy. The computational cost of DD, however, is considerably big, because it has to compute all distances between all combinations of instances. This report presents the approximate computation of DD for MIIR using nearest neighbor search. We confirmed through the experiments that the computational speed of DD becomes 520 times faster on Nister's database while keeping the accuracy.

Gaussian process regression is a nonlinear regression that estimates the expected value of the of the output and its variance for given input. The original Gaussian process regression estimates a scalar value as an expected value of output, which can simply be extended to estimate a vector value. However, the covariance matrix cannot be estimated by simple extension. We have proposed an accelerated Gaussian process regression by introducing dynamic active set consisting of input-output pairs, and the weighted output covariance can be utilized as the covariance of output for multivariate Gaussian process regression. However, the diagonal elements of the estimated matrix can be negative. This is caused by the negative weight of the outputs originated by the inverse of gram matrix. In this report, we propose a method to estimate non-negative weights for the outputs. The estimation is twofold: initially estimate the output vector by simple multivariate Gaussian process regression, then recompute non-negative weights so as to minimize the output error. The resulted weights are utilized to estimate covariance matrix. By using the proposed method in anomaly detection of plant data and electrocardiogram data in the experiment, it was confirmed its effectiveness.

This report presents a cooperation method for multiple pedestrian detection systems, each of which has a single camera with wide view angle, for extending their observation area. Most single-viewpoint human detection algorithms perform matching between pedestrian models and captured images. Since the major task of the target system is people counting, the output of the system is rather simple. We assume that only the pedestrian positions on the image plane are available. For the cooperation of such systems, we propose map-based cooperation scheme, where the detected positions by multiple systems are transformed by nonlinear regression onto a single floor map, and the locations detected by different systems are merged based on the proximi ty. In our research, we employ linear regression tree for the nonlinear regression due to its short computation time. However, single camera people detection system cannot produce accurate floor position without the height information of each pedestrian. This is because the half upper body of the human model has strong matching weight, and hence, the head position is well estimated on the image plane but the foot position is not. Our method computes the correspondences of the locations of the same pedestrian, and the pedestrian's height is estimated simultaneously based on the discrepancy of their locations on the map.

For Multiple Instance Image Retrieval (MIIR), we have proposed index feature reduction method, which reduces the number of indexes attached to the image entries in the database. This method computes the importance measure representing the stability and the discrimination power of each feature by using the framework of Diverse Density and reduces the features having less importance measures. Through the experiments, this reduction drastically reduces the memory usage and retrieval accuracy, but the acceleration of retrieval speed is limited. This is because the number of nearest neighbor searches performed for each query is equivalent to the number of local features in the query image. This means that query feature reduction is required for the acceleration of image retrieval. This report presents a query feature reduction method for solving this problem.

Content Based Image Retrieval (CBIR) using local features can be classified into two types. One is the method using single vector representation that integrates local features detected in an image into a single Bag of Feature (BoF) vector. The other uses multiple instance representation without integration. We call the latter method Multiple Instance Image Retrieval (MIIR). In MIIR, a method for reducing database indexes has been proposed. This method employs the framework of Diverse Density (DD) to represent the importance measure, which means the stability as well as the discriminative power of the feature (instance). This reduction reduces the memory usage and the retrieval accuracy. The computational cost of DD, however, is considerably big, because it has to compute all distances between all combinations of instances. This report presents the approximate computation of DD for MIIR using nearest neighbor search. We confirmed through the experiments that the computational speed of DD becomes 520 times faster on Nister's database while keeping the accuracy.

Object re-identification techniques are essential to improve the identification performance in video surveillance tasks. The re-identification problem is equal to a multi-view problem that an unknown individual is identified across spatially disjoint data. For the re-identification techniques, several multi-view feature transformation methods have been proposed. These methods are formulated by the supervised learning framework and show the better performances in multi-view classification tasks in which the training data are observed by the different sensors. However, in the re-identification tasks, these methods may not be required because the simple feature transformation method such as linear discriminant analysis (LDA) shows the reasonable identification rates. In this paper, we propose a novel semi-supervised feature transformation method, which is formulated as a natural coupling with PCA and LDA modeled by the graph embedding framework. Our method showed best re-identification performances compared with other feature transformation methods.

Face image recognition is widely used for many applications from crime investigation to user identification for personal devices. Most face recognition systems, however, do not show the reason of recognition, i.e., by which part of the face, this image is recognized as the resulted person. This paper proposes a face image recognition method that provides the recognition result, as well as the reason of the result. This method is basically an image recall system using example based nonlinear mapping. This nonlinear mapping refers the image database and performs image-to-image mapping using the database so as to realize image recall, where the system generates the similar image with the input by partially linear combination of images in the database. As a by-product, weighting coefficients for images in the database are obtained at any locations on the recalled image. By applying this method to face images, we can recognize the person id by summing up the coefficients for ids. The coefficient distribution for the resulted person id, we can understand which part provides the evidence of the recognition result. Through experiments conducted on ORDB face database, we confirmed our method have high recognition rate.

Crime investigation based on surveillance videos requires time and manpower consuming work, where the main task is to find the same person captured by different videos. This task is important for tracing the suspects or related people, and the results sometimes provide important cues for solution. However, so many surveillance videos are working in urban areas, and the investigators gather them and inspect them by using many workers for long time. If this task is automated, we can reduce the number of workers and overlooking. Also the automated system works faster than human inspection, we can accelerate the investigation process. This paper proposes a system finding the image pair capturing the same person based on local features. This implies our method does not require human detection or tracking. This property is suitable for practical situations, where people are occluded by obstacles, and most human detection systems fail. Our method detects local features for input image and by suppressing those features detected at the background region. The resulted local features are assigned as positive images and by computing the Commonality measure derived from Diverse Density, we can find the image pairs having high commonality. We applied our method to multiple video sequences, and confirmed that the proposed method is promising when the local features are roughly preserved.

Crime investigation based on surveillance videos requires time and manpower consuming work, where the main task is to find the same person captured by different videos. This task is important for tracing the suspects or related people, and the results sometimes provide important cues for solution. However, so many surveillance videos are working in urban areas, and the investigators gather them and inspect them by using many workers for long time. If this task is automated, we can reduce the number of workers and overlooking. Also the automated system works faster than human inspection, we can accelerate the investigation process. This paper proposes a system finding the image pair capturing the same person based on local features. This implies our method does not require human detection or tracking. This property is suitable for practical situations, where people are occluded by obstacles, and most human detection systems fail. Our method detects local features for input image and by suppressing those features detected at the background region. The resulted local features are assigned as positive images and by computing the Commonality measure derived from Diverse Density, we can find the image pairs having high commonality. We applied our method to multiple video sequences, and confirmed that the proposed method is promising when the local features are roughly preserved.

Appearance based 3D object modeling is a handy approach for 3D object recognition. The most representative method in this approach is parametric-eigenspace method, which models the appearance variations of a 3D object by a manifold in eigenspace and recognizes the object name and the observation direction by using the manifolds of multiple objects. This method, however, is not robust against occlusions and background variations. For solving this problem, this paper proposes a robust 3D object modeling and recognition method based on local features. The inputs of this method are the images of objects representing their aspects. From these images, tons of local features are generated. Some are unique to an object and some are not. Also, some are created by specular highlight and shadows. By grouping the appearances of contiguous aspects, we can obtain a compact representation of an object, which indirectly excludes fragile and non-essential features. However, excessive grouping may cause poor recognition rate. The distinctive features unique to an object can be found by using the framework of Diverse Density (DD). The DD can be used also for controlling the grouping process. Our DD guided local feature selection and grouping is applied to Coil image dataset for object recognition task. Through comparative experiments with parametric-eigenspace method for occluded and background added input images, we confirmed that our method achieves much higher recognition rate.

Appearance based 3D object modeling is a handy approach for 3D object recognition. The most representative method in this approach is parametric-eigenspace method, which models the appearance variations of a 3D object by a manifold in eigenspace and recognizes the object name and the observation direction by using the manifolds of multiple objects. This method, however, is not robust against occlusions and background variations. For solving this problem, this paper proposes a robust 3D object modeling and recognition method based on local features. The inputs of this method are the images of objects representing their aspects. From these images, tons of local features are generated. Some are unique to an object and some are not. Also, some are created by specular highlight and shadows. By grouping the appearances of contiguous aspects, we can obtain a compact representation of an object, which indirectly excludes fragile and non-essential features. However, excessive grouping may cause poor recognition rate. The distinctive features unique to an object can be found by using the framework of Diverse Density (DD). The DD can be used also for controlling the grouping process. Our DD guided local feature selection and grouping is applied to Coil image dataset for object recognition task. Through comparative experiments with parametric-eigenspace method for occluded and background added input images, we confirmed that our method achieves much higher recognition rate.

Face image recognition is widely used for many applications from crime investigation to user identification for personal devices. Most face recognition systems, however, do not show the reason of recognition, i.e., by which part of the face, this image is recognized as the resulted person. This paper proposes a face image recognition method that provides the recognition result, as well as the reason of the result. This method is basically an image recall system using example based nonlinear mapping. This nonlinear mapping refers the image database and performs image-to-image mapping using the database so as to realize image recall, where the system generates the similar image with the input by partially linear combination of images in the database. As a by-product, weighting coefficients for images in the database are obtained at any locations on the recalled image. By applying this method to face images, we can recognize the person id by summing up the coefficients for ids. The coefficient distribution for the resulted person id, we can understand which part provides the evidence of the recognition result. Through experiments conducted on ORDB face database, we confirmed our method have high recognition rate.

Discriminating feature extraction is important to achieve high recognition rate in a classification problem. Fisher's linear discriminant analysis (LDA) is one of the well-known discriminating feature extraction methods and is closely related to the Mahalanobis distance metric learning. Neighborhood component analysis (NCA) is one of the Mahalanobis distance metric learning methods based on stochastic nearest neighbor assignment. The objective function of NCA can be expressed as a within-class coherency by a simple formula, and NCA extracts discriminating features by minimizing the objective function. Unfortunately, the computational cost of NCA significantly increases as the number of input data increases. For reducing the computational cost, we propose a fast distance metric learning method by taking the between-class distinguishability into account of nearest mean classification. According to the experimental results using standard repository datasets, the computational time of our method is evaluated as 27 times shorter than that of NCA while keeping or improving the accuracy.

This report presents a method for estimating color face images from near-infrared monochrome face images. This estimation is done by the regression from a monochrome image to a color image. One difficult problem is that the regression depends on face organs. That is, the same intensity pixels in an infrared monochrome image do not correspond to the same color pixels. Therefore, entirely uniform regression cannot colorize the pixels correctly. This report presents a colorization method for monochrome face images by position-dependent regressions, where the regression coefficients are obtained in different image regions corresponding to facial organs. Also, we can extend the independent variables by adding texture information around the pixels so as to obtain accurate color images. However, unrestricted extension may cause multi-collinearity problem, which may produce inaccurate results. This report also proposes CCA based dimensionality reduction for avoiding this problem. Comparative experiments on the restoration accuracy demonstrate the superiority of our method.This report presents a method for estimating color face images from near-infrared monochrome face images. This estimation is done by the regression from a monochrome image to a color image. One difficult problem is that the regression depends on face organs. That is, the same intensity pixels in an infrared monochrome image do not correspond to the same color pixels. Therefore, entirely uniform regression cannot colorize the pixels correctly. This report presents a colorization method for monochrome face images by position-dependent regressions, where the regression coefficients are obtained in different image regions corresponding to facial organs. Also, we can extend the independent variables by adding texture information around the pixels so as to obtain accurate color images. However, unrestricted extension may cause multi-collinearity problem, which may produce inaccurate results. This report also proposes CCA based dimensionality reduction for avoiding this problem. Comparative experiments on the restoration accuracy demonstrate the superiority of our method.

Face image retrieval based on local features has advantages of short elapsed time and robustness against the occlusions. However, the keypoint detection, beforehand with the feature description, may fail due to illumination changes. For solving this problem, top-down model-based keypoint detection can be applied, where man-made face model does not fit this task. This report addresses the problem of bottom-up face model construction from example, which can be formalized as common local features extraction among face images. For this purpose, a measure called Diverse Density (DD) can be applied. DD at a point in a feature space represents how the point is close to other positive example while keeping enough distance from negative examples. Because of this property, DD is defined as product of metrics, which can easily be affected by exceptional data, i.e., if one negative data leaps into the neighbour of a positive example, the DD around there becomes lower. Actually, face images have wide variations of face organs' positions, beard, moustache, glasses, and so on. Under these variations, DD for wide varieties of face images will be low at any point in the feature space. For solving this problem, we propose a method performing hierarchical clustering and common local feature extraction simultaneously. In this method, we define a measure representing the affinity of two face image sets, and cluster the face images by iteratively merging the cluster pair having the maximum score. Through experiments on 1021 CAS-PEAL face images, we confirmed that multiple face models are successfully constructed.Face image retrieval based on local features has advantages of short elapsed time and robustness against the occlusions. However, the keypoint detection, beforehand with the feature description, may fail due to illumination changes. For solving this problem, top-down model-based keypoint detection can be applied, where man-made face model does not fit this task. This report addresses the problem of bottom-up face model construction from example, which can be formalized as common local features extraction among face images. For this purpose, a measure called Diverse Density (DD) can be applied. DD at a point in a feature space represents how the point is close to other positive example while keeping enough distance from negative examples. Because of this property, DD is defined as product of metrics, which can easily be affected by exceptional data, i.e., if one negative data leaps into the neighbour of a positive example, the DD around there becomes lower. Actually, face images have wide variations of face organs' positions, beard, moustache, glasses, and so on. Under these variations, DD for wide varieties of face images will be low at any point in the feature space. For solving this problem, we propose a method performing hierarchical clustering and common local feature extraction simultaneously. In this method, we define a measure representing the affinity of two face image sets, and cluster the face images by iteratively merging the cluster pair having the maximum score. Through experiments on 1021 CAS-PEAL face images, we confirmed that multiple face models are successfully constructed.

This paper presents a show-through removal method based on discrete optimization. Images of duplex printed documents may be degraded by show-through, which may affect the character recognition ratio of a document. Show-through is a common phenomenon that texts or figures on one side appears on the other side. There are two ways to remove show-through components from the images: show-through removal from both-side images and that from a one-side image. The method using both-side images requires 1)both side images and 2)precise alignment of images, which is a difficult problem to solve automatically. The method using a one-side image can be designed as a discrete optimization problem estimating both-side intensities for each pixel. However, the solution requires huge computational cost, because the number of labels is big (256×256). In this report, we formalize this problem as another discrete-optimization problem to estimate pixel's attributes (foreground, show-through, or background). Clearly, this method uses three labels, and thus, computationally inexpensive. Through experiment using many show-through images, we confirmed that our method removes show-through components.

This report presents an "external" k-means clustering on HDD K-means clustering is widely used for many applications. For example, codebook creation for Bag of Visual Words requires k-means clustering on huge amount of local feature vectors to obtain Visual Words (codebook entries). Standard "internal" k-means clustering loads the whole vector data on the main memory and performs clustering. This working memory can explode for huge amount of data. As a solution of this problem, we propose an "external" clustering algorithm on HDD. This is a multi-path algorithm, which scans the whole data in each path In the first stage, cluster centroids are updated gradually, providing the data sequentially. Through this path, the number and the sum of the data are recorded for each cluster, and the belonging cluster is recorded for each data. In the following paths, each data is provided and the cluster center is updated for those data that changes belonging cluster. By adjusting this update frequency, the number of distance computation can be reduced and the performance can be improved.

This report presents a nearest neighbor search algorithm without distance computation K-D tree based nearest neighbor search algorithm consists of two processes : 1) determine tentative nearest neighbor data by tree search, and 2) A* algorithm based "priority search" to find true nearest neighbor data to the query. The latter process consumes most computational time especially for high dimensional data distributions. We focus on the fact that the nearest neighbor data is essentially determined by the query position in the feature space and the priority search can be replaced by classifiers. In practice, we employ decision trees as classifiers and attach them to the K-D tree leaf nodes. This data structure is named Grafting trees. On this tree, we can find the nearest neighbor data without distance computation, In this data structure, some decision tree leaves can be very deep. For solving this problem, we modified this data structure forgiving distance computation for such leaf nodes. The modified tree is named Hybrid Grafting trees. In the experiments, we compared ANN library with our methods and confirmed the search time is shorter than ANN.

Face image retrieval based on local features has advantages of short elapsed time and robustness against the occlusions. However, the keypoint detection, beforehand with the feature descnption, may fail due to illumination change. For solving this problem, top-down model-based keypoint detection must be effective, where man-made face model does not fit for this task. This report addresses the problem of bottom-up face model creation from examples, which can be formalized as common local feature extraction among examples. For this purpose, a measure called Diverse Density (DD) established in the field of Multiple Instance Learning (MIL) can be applied. DD at a point in a feature space represents how the point is close to other positive examples while keeping enough distance from negative examples. Because of this this property, DD is defined as a product of metrics, which can easily be affected by exceptional data, i.e., if one negative data leaps into the neighbor of a positive example, the DD around there becomes lower. Actually, face images have wide vanations of face organs' positions, beard, mustache, glasses, and so on. Under these vanations, DD for wide varieties of face images will be low at any point in the feature point. For solving this problem, we propose a method performing hierarchical clustering and common feature extraction simultaneously. In this method, DD score is employed as a measure representing the integrity of the face image set, and hierarchical clustering is performed by merging the cluster pair having maximum DD score. Through experiments on 1021 CASPEAL face images, we confirmed that multiple face models are successfully constructed.

Appearance learning of an object represented by a text can be realized by utilizing image retrieval systems on the Internet. This enables us a real-world object search system, which searches real object providing a text representing the object. Images collected by an image retneval system, however, cannot be used for appearance learning, because the object locations and sizes are not uniform among collected images. That is, object localization in each image is required. This prob-lem can be taken as a co-segmentation problem extracting common object among images, and we propose a method based on Multiple-Instance Learning (MIL) framework. Our method consists of two parts, foreground-background modeling and dis-crete optimization to obtain object regions. Foreground-background modeling is to compute Diverse Density (DD) for exces-sively partitioned image regions, where original DD represents similanty among positive features and dissimilarity between negative features. In our method, we slightly modified the DD definition to fit the co-segmentation problem Expenmental results using iCoseg database demonstrates higher accuracy of our method than other methods.

This paper presents an active tracking system that only uses depth image taken by Kinect. From the view pomt of Kinect, flexible software libranes including human detection and human-pose estimation cannot be applied when target people go outside of the viewing area, and the active trackmg system solve or relax this problem. From the active-tracking -system viewpoint, different from other cameras, Kmect can take depth image in mdoor scene independent of illumination conditions. This implies that the tracking system can keep tracking even in complete darkness. For single camera trackmg system, Fixed Viewpoint Camera (FVC) setting, which does not change the 3D viewpoint under rotation, is useful for active tracking, because the rotation angle to keep target inside the image frame can directly computed from the 2D image position on the image plane. In the case of Kmect, however, we don't have to use FVC settings if we know the relative viewpoint position from the rotation center, because 3D positional information is directly obtained from the sensor. First, we present the calibration method to know the relative position. Next, we present a 3D blob-tracking method based on mean-shift Since this method uses 3D window function, the object separability from its background and the stability are both good when the target shape is roughly convex. As a tracking method using more precise 3D model, this paper also presents spin-image based tracking algorithm. Through the expenment, we discuss what method should be used for specific tasks.

This report presents a method to estimate vector outputs m the framework of Gaussian Process Regression (GPR). GPR is a non-linear regression method based on input-output examples. Since basic GPR estimates an output consisting of scalar mean and variance, multiple executions of GPR cannot estimate covariance between vector components. Our method estimates both vector mean values and covariance matrices based on our previous method Dynamic Active Set (DAS). Active set is the set of examples consisting of input and output pairs. DAS automatically construct an active set suitable for estimating the output for given mput. Our method estimates covariance matrix from the output components of the. Active Set obtained by DAS Experimental results for artificial and real datasets demonstrate the soundness of our method.

There exists a calibration method that modifies depth and color images captured by Kinect so as to minimize the 2D geometric inconsistencies. However, the resulting image pairs still have small 2D allocation error. Furthermore, the distance measurement may fail due to weak reflections at dark area, specular reflections, infra-red light occlusion observed near occluding contour, and so on. Under these geometric inconsistency and lack of depth information, re-projection of the color image projected depth map to different screen may produce heavily distorted image. This report presents a depth image correction method using local relationship between depth and color information, i.e., spatially near and similar-color pixels should have similar depth. For this computation, bilateral filtering best fits, where the product of color-similarity and spatial proximity represents the weight of the depth information. Through some experiments, we confirmed that our method produces accurately allocated and lack free depth images.

Phase-shift 3-D measurement is a very accurate depth measurement method that can measure the depth in sub micrometer order by analyzing at least 3 images. The resolution is very high, however, the depth is initially estimated as a phase value in a certain depth period, which corresponds to a depth interval. For estimating the absolute depth from the phase information, we have to determine the order of the depth period. This problem is called "phase unwarping" problem. Popular solution of phase unwarping requires additional light projections and observations, which reduce the temporal resolution of the depth measurement. This report formalizes the phase unwarpmg as a discrete optimization problem. The problem is solved by the belief-propagation. Our previous method only uses smoothness-term in the formalization, however, this report presents a method to use data-term at some pixels by modifying the projection light pattern. Through experiments, we confirmed that the depth reconstruction is successfully done for a "steep" object whose depth spatially changes within a small position change.

This report presents a method that estimates a color face image from a monochrome face image taken under near-infrared illumination. The estimation is done by the regression between parameters obtained by CCA, where the regression coefficients are learned from color and infrared image pairs a pnon. The most difficult problem is that the regressions depend on face organs. That is, the same intensity pixels in an infrared monochrome image do not correspond to the same color pixels in a color image. That is, globally uniform regression cannot colorize the pixels correctly. Especially, intensities of iris region in an infrared image are brighter than that in the corresponding monochrome image taken under visible ray illumination. In our method, we colonze a monochrome face image by regressions, where the coefficients are obtained in different image regions corresponding to facial parts. These regions are obtained by two-stage segmentation, which uses learned face template and adaptive segmentation. Comparative experiments between simple intensity based segmentation and our two-stage segmentation demonstrates the restoration accuracy of our method.

We are planning to construct an image retrieval system using FPGA. For designing this system, we developed a prototype system, which retrieves the image having maximum number of matched local image features with a query image. The reason why we don't use Bag of Features (BoF) is that codebook referencing may consume considerable time and computational resources on FPGA. For this purpose, the local features describing a stored image should satisfy the following conditions : 1) they should have strong discrimination power from other images, 2) they should be robust agamst observation distortions including rotation, scaling, and so on. In order to maximize the number of stored images, the number of local features describing a stored image should be minimized For selecting such "good local features" from all local features, we propose a method based on Diverse Density. In the experiment, we confine the number of local features describing a single image to 10, and our method outperforms other local feature selection methods.

インターネット上の画像検索を利用して，物体の見えの学習を行うことができれば，テキストを与えるだけで現実世界での物体の探索が行えるようになる．しかし，現実の画像検索結果は，対象以外の無関係なものを含むため，画像内での対象のローカライズが必要になる．本報告では，この問題を，複数の画像群から共通対象を検出するCo-segmentationの問題であると捉え，Multiple-InstanceLearning(MIL)の考え方を取り入れて，この問題を解く方法を提案する.具体的には,過剰分割した画像領域から抽出した特徴量に対して求めたDiverseDensity(DD)の値を利用して前景・背景モデルを構築し，離散最適化によって，各画像で共通する領域の検出を行う．この際に通常のDDを計算したのでは，前景におけるDDの値が低下するために，Co-segmentationに適した新たな尺度を導入する．iCoseg databaseを用いた実験では先行研究に比べ高いセグメンテーション精度が得られ，本手法の有効性が確認できた．Appearance learning of an object represented by a text can be realized by utilizing image retrieval systems on the Internet. This enables us a real-world object search system, which searches real object providing a text representing the object. Images collected by an image retrieval system, however, cannot be used for appearance learning, because the object locations and sizes are not uniform among collected images. That is, object localization in each image is required. This problem can be taken as a co-segmentation problem extracting common object among images, and we propose a method based on Multiple-Instance Learning (MIL) framework. Our method consists of two parts, foreground-background modeling and dis crete optimization to obtain object regions. Foreground-background modeling is to compute Diverse Density (DD) for exces sively partitioned image regions, where original DD represents similarity among positive features and dissimilarity between negative features. In our method, we slightly modified the DD definition to fit the co-segmentation problem. Experimental results using iCoseg database demonstrates higher accuracy of our method than other methods.

位相シフト法は，最低３枚の画像を解析するだけで，画像全体の三次元形状をサブ川単位の分解能で計測できる手法である．しかしながら，各画素で求められる位相は，奥行きを幾つかの周期に区分した内部での奥行きに対応付けられるため，各位相情報が何周期目の区分の中での奥行きかを求めない限り，一意に奥行きを決定できない．これを補うために，各位相が何周期目の位相かを決定する「位相接続」問題を解かなければならない．本論文では，この問題を離散最適化問題として定式化し，BeliefPropagationを用いて解くことにより，奥行きの推定を行う．特に，以前は相対的な形状を求めるために，平滑化項のみを用いた離散最適化を行なってきたが[1]，今回は投光パターンを工夫することにより，幾つかの点で絶対的奥行きを表すデータ項を求め，これを用いて奥行き推定を行う手法を提案する．実験により，奥行きが激しい物体であっても正しく奥行きの推定が行え，提案手法の有効性を確認した．Phase-shift 3-D measurement is a very accurate depth measurement method that can measure the depth in submicrometer order by analyzing at least 3 images. The resolution is very high, however, the depth is initially estimated as a phase value in a certain depth period, which corresponds to a depth interval. For estimating the absolute depth from the phase information, we have to determine the order of the depth period. This problem is called "phase unwarping" problem. Popular solution of phase unwarping requires additional light projections and observations, which reduce the temporal resolution of the depth measurement. This report formalizes the phase unwarping as a discrete optimization problem. The problem is solved by the belief-propagation. Our previous method only uses smoothness-term in the formalization, however, this report presents a method to use data-term at some pixels by modifying the projection light pattern. Through experiments, we confirmed that the depth reconstruction is successfully done for a "steep" object whose depth spatially changes within a small position change.

我々は，SIFTやSURFなどの画像の局所特徴量を用いた画像検索システムをFPGA上に構築するためのプロトタイプを設計している．この設計ではBag of Features(BoF)などのコードブックを用いた特徴記述を行うことは煩雑であるため，局所特徴量そのものを用いた画像検索を行う．この際,一枚の画像を特徴付けるキーポイントが満たすべき条件としては，１）他の画像には含まれない弁別性の高い局所特徴量を持つこと，２）画像の回転やスケール変化等の変換を受けても検出されやすいこと，という２つが重要である．これらの条件を満足する少数のキーポイントを用いて登録する画像を記述しておけば，省メモリかつ高速で，しかもロバストな画像検索が行えるはずである．このようなキーポイントを抽出するためにDiverse Densityを用いたキーポイントの絞込み法を提案する．実験では，１枚の画像を１０個のキーポイントで記述するものとし，ランダムに選んだキーポイントよりも，提案手法で選んだ１０点のほうが圧倒的に安定な検索が行えることを確認した．We are planning to construct an image retrieval system using FPGA. For designing this system, we developed a prototype system, which retrieves the image having maximum number of matched local image features with a query image.The reason why we don't use Bag of Features (BoF) is that codebook referencing may consume considerable time and computational resources on FPGA. For this purpose, the local features describing a stored image should satisfy the following conditions: 1) they should have strong discrimination power from other images, 2) they should be robust against observation distortions including rotation, scaling, and so on. In order to maximize the number of stored images, the number of local features describing a stored image should be minimized. For selecting such "good local features" from all local features, we propose a method based on Diverse Density. In the experiment, we confine the number of local features describing a singleimage to 10, and our method outperforms other local feature selection methods.

本報告では，GaussianProcessRegression(GPR)の枠組みを用いてベクトル値の出力を推定する手法を提案する．ＧＰＲは出力の期待値と分散を推定する非線形回帰手法である．基本的なＧＰＲで推定できる出力の期待値はスカラー値であり,ベクトル値を推定する際にＧＰＲを出力次元の回数だけ実行しても各出力次元間の共分散は推定できない．本手法では，我々が先に提案したActiveSetの動的紋込みを利用し，ベクトル出力の期待値とその共分散の推定を行う．ActiveSetの動的絞込みは，入力データをキーとして，出力の推定に必要な事例集合を動的に決定する方法である．各事例は，入力と出力のペアから構成されており，ActiveSetの出力部を利用することで，共分散行列の推定を行うことが出来る．実験では，提案手法を人工的なデータとプラントデータに適用し，提案手法の有効性を確認した．This report presents a method to estimate vector outputs in the framework of Gaussian Process Regression(GPR) . GPR is a non-linear regression method based on input-output examples. Since basic GPR estimates an output consisting of scalar mean and variance, multiple executions of GPR cannot estimate covariance between vector components.Our method estimates both vector mean values and covariance matrices based on our previous method Dynamic Active Set (DAS). Active set is the set of examples consisting of input and output pairs. DAS automatically construct an active set suitable for estimating the output for given input. Our method estimates covariance matrix from the output components of the Active Set obtained by DAS. Experimental results for artificial and real datasets demonstrate the soundness of our method.

局所特徴量に基づく顔画像検索は，高速でありオクルージョンにも頑健であるが，照明変化によってキーポイントの検出漏れが発生するため，特徴記述が不安定になる．この問題に対処するには，モデルを用いたトップダウン的なキーポイント検出が有効であると考えられる．しかし，人間の先見的知識に基づいてこれらのモデルを構築した場合，必ずしも実際の画像とは一致しないため，実際のデータからボトムアップ的に画像の局所特徴量に基づくモデルを構築しなければならない画像の局所特徴量の共通性を評価する尺度として，Multiple lnstanceLeaming(MIL）で用いられるDiverse Density(DD)がある．これを用いれば，非顔画像から抽出したネガティブデータから遠く，顔画像から抽出したポジティブデータと近い局所特徴量が抽出できる．しかし，実際の人物顔画像は顔器官の空間的配置や，メガネや髭など，変化に富み，全ての顔画像について積の形で表現されるDDの値を計算したのでは値が小さくなり，共通性の高い顔特徴は検出できない．本報告では，この問題を解決するために，DDの期待値をスコアとして，顔画像集合を階層的にクラスタリングしながら，局所特徴量のモデルを構築する方法を提案する．実験では，CASPEALの顔画像1021枚を対象として提案手法を実行し，複数の顔モデルの抽出が行えることを確認した．Face image retrieval based on local features has advantages of short elapsed time and robustness against the occlusions. However, the keypoint detection, beforehand with the feature description, may fail due to illumination change. For solving this problem, top-down model-based keypoint detection must be effective, where man-made face model does not fit for this task. This report addresses the problem of bottom-up face model creation from examples, which can be formalized as common local feature extraction among examples. For this purpose, a measure called Diverse Density (DD) established in the field of Multiple Instance Learning (MIL) can be applied. DD at a point in a feature space represents how the point is close to other positive examples while keeping enough distance from negative examples. Because of this this property, DD is defined as a product of metrics, which can easily be affected by exceptional data, i.e., if one negative data leaps into the neighbor of apositive example, the DD around there becomes lower. Actually, face images have wide variations of face organs' positions,beard, mustache, glasses, and so on. Under these variations, DD for wide varieties of face images will be low at any point in the feature point. For solving this problem, we propose a method performing hierarchical clustering and common feature extraction simultaneously. In this method, DD score is employed as a measure representing the integrity of the face image set,and hierarchical clustering is performed by merging the cluster pair having maximum DD score. Through experiments on 1021CASPEAL face images, we confirmed that multiple face models are successfully constructed.

本報告では，Kinectによって得られる深度画像のみを用いて，首振りによる対象の能動追跡を行うシステムおよび方法について述べる．これは，Kinectの側から見ると，人物検出やポーズ推定などを高速かつ安定に行うソフトウエアライブラリが使えるにもかかわらず，人物が観測範囲から出てしまうとその機能が使えなくなることを補うことになる．また，能動追跡の立場から見ると，屋内であれば照明の強弱にかかわらず，完全な暗闇でも追跡が行えることを意味している．単眼およびステレオカメラなどでは，回転による運動視差が発生しない「視点固定型」のカメラを用いた能動追跡が２次元画像平面上での対象位置のみで回転角が決定できるため有利である．しかし，Kinectでは３次元情報が直接得られるため，深度カメラの光学中心と回転中心の位置関係が分かっていれば，視点固定型の構成に剃る必要はない本稿では，まずこのキャリブレーション法について述べる．次に，３次元空間中での点群をmeanshiftによって追跡する手法について述べる．この手法は，３次元の窓関数を用いるため，物体が凸であり周辺に他の物体が存在しない限り，安定な追跡を実現することが出来る．次に，より詳細な３次元形状に基づく追跡アルゴリズムの例としてspmimageを用いた追跡方法，Ｐ、制御の方法について述べ，実験結果から，対象形状やタスクによって，どのような追跡手法が適しているかについて述べる．This paper presents an active tracking system that only uses depth image taken by Kinect. From the view point of Kinect,flexible software libraries including human detection and human-pose estimation cannot be applied when target people go outside of the viewing area, and the active tracking system solve or relax this problem. From the active-tracking -system viewpoint, different from other cameras, Kinect can take depth image in indoor scene independent of illumination conditions. This implies that the tracking system can keep tracking even in complete darkness. For single camera tracking system, Fixed Viewpoint Camera (FVC) setting, which does not change the 3D viewpoint under rotation, is useful for active tracking, because the rotation angle to keep target inside the image frame can directly computed from the 2D image position on the image plane. In the case of Kinect, however, we don't have to use FVC settings if we know the relative viewpoint position from the rotation center, because 3D positional information is directly obtained from the sensor. First, we present the calibration method to know the relative position. Next, we present a 3D blob-tracking method based on mean-shift. Since this method uses 3D window function, the object separability from its background and the stability are both good when the target shape is roughly convex. As a tracking method using more precise 3D model, this paper also presents spin-image based tracking algorithm. Through the experiment,we discuss what method should be used for specific tasks.

This paper proposes a method for detecting pedestrians on the road in night view using HOG, NFTG and the geometric constraints that pedestrians stand or walk on the road. The vanishing point of the road is estimated from the white lines on the road using particle filter and Hough transform. The vanishing point is used to predict the size of the pedestrian when whose position is given, thus brings the effect of reducing the processing time and the false detection. A new method called "Integral-Ni-HOG" (or INi-HOG) for calculating HOG for near-infrared images fast is proposed. The Hausdorff distance between the pedestrian candidate regions and the pedestrian model is used for the verification of each pedestrian candidate region in the image. NFTG is used to describe the relations between each pedestrian model in order to reduce the number of times of pedestrian models being used for verifying each pedestrian candidate region. The experiments using real images show that this method is 15 times faster and the false detection rate is lowered by more than 50% compared the conventional method based on HOG.

本稿では幾何制約付 INi-HOG と NFTG (Nearest First Traversing Graph) を使って夜間歩行者を検出する方法を提案する．歩行者候補領域の検出段階では、安定に消失線を求めるために、路面内の白線等の直線情報に基づいて，ハフ変換とパーティクルフィルタから消失点を検出する方法 [15] を導入する．近赤外線画像用の HOG(Ni-HOG)[14] を高速に計算するために、Integral-Ni-HOG（INi-HOG） を提案する．消失線とカメラパラメータから INi-HOG の探索位置とサイズを導出でき，高速処理かつ誤検出が抑制することが可能になる．歩行者候補領域の検証段階では、人物輪郭モデルとのハウスドルフ距離を計算する方法 [14] を採用する．多数の人物輪郭モデルとの距離計算回数を最小限にするために、NFTG[7] を用いたモデル間のグラフ化を提案する。実画像を用いた実験結果より，従来手法 (HOG) と先行研究 [14] と比較し，提案手法を用いた場合，誤検出を約半分まで抑えながら，処理速度が約 15 倍速くなることを確認した．This paper proposes a method for detecting pedestrians on the road in night view using HOG, NFTG and the geometric constraints that pedestrians stand or walk on the road. The vanishing point of the road is estimated from the white lines on the road using particle filter and Hough transform[15]. The vanishing point is used to predict the size of the pedestrian when whose position is given, thus brings the effect of reducing the processing time and the false detection. A new method called "Integral-Ni-HOG" (or INi-HOG) for calculating HOG for near-infrared images fast is proposed. The Hausdorff distance between the pedestrian candidate regions and the pedestrian model[14] is used for the verification of each pedestrian candidate region in the image. NFTG[7] is used to describe the relations between each pedestrian model in order to reduce the number of times of pedestrian models being used for verifying each pedestrian candidate region. The experiments using real images show that this method is 15 times faster and the false detection rate is lowered by more than 50% compared the conventional method[14] based on HOG.

我々は，プラントに取り付けられたセンサ情報から，プラントの異常や故障等の予兆を検出する方法として，事例ベースの内挿計算法である Gaussian Processes を用いた手法を提案している．Gaussian Processes は，出力の期待値だけでなく，標準偏差も同時に推定することが出来るという特長を有している．しかし，センサ信号の異常さは，瞬時的に現れる場合と，比較的長期間の趨勢として現れる場合がある．長期間の分析のために入力ベクトルの次元数を増加させると，精度を保つために必要な事例数が急激に増大するため，可用性が著しく低下する．この問題に対処するために，時間軸方向に可変窓幅で信号を平滑化する多重時間解像度解析を導入した異常予兆検出法を提案する．この手法により，様々なスケールで起きる異常の予兆をより確実に検出できるようになる．We have been proposed a pre-fault detection system based on Gaussian Processes (GP). The advantage of GP based pre-fault detection is that it can estimate not only the estimated output but also the standard deviation of the output. However, the anomaly can sometimes be detected as short-term phenomena and sometimes as long-term tendency. If we simply apply GP to high-dimensional vectors sampled within wide temporal window, the method requires huge number of training samples to keep the sensitivity and the accuracy. This is because GP is an example based non-linear regression method. For solving this problem, we introduce a multi-resolutional analysis along temporal axis that produces multiple signals with different smoothing windows from a single signal. By applying our GP based pre-fault detection method to these multi-scale signals, we can realize a versatile and reliable pre-fault detection system.

We have been proposed a pre-fault detection system based on Gaussian Processes (GP). The advantage of GP based pre-fault detection is that it can estimate not only the estimated output but also the standard deviation of the output. However, the anomaly can sometimes be detected as short-term phenomena and sometimes as long-term tendency. If we simply apply GP to high-dimensional vectors sampled within wide temporal window, the method requires huge number of training samples to keep the sensitivity and the accuracy. This is because GP is an example based non-linear regression method. For solving this problem, we introduce a multi-resolutional analysis along temporal axis that produces multiple signals with different smoothing windows from a single signal. By applying our GP based pre-fault detection method to these multi-scale signals, we can realize a versatile and reliable pre-fault detection system.

位相シフト法は，比較的少数の画像の解析のみで高精度な形状計測を行うことが可能である．しかしながら，投影した格子パターンの位相が何番目の周期に含まれるかを求めないと物体全体の三次元形状の計測を行えない問題がある．本論文では，この問題を離散最適化問題として定式化し，Belief Propagation（信念伝搬）法を用いて解くことにより，位相接続を行う方法を提案する．本手法では，計測物体の絶対的な奥行きではなく，各点の周期を決定することで滑らかな三次元形状を計測することを目的にしている．

本論文では，インターネットの画像検索を用いて集められた画像群において，個々の画像上に写る対象の領域を推定する手法を提案する．提案手法では，複数の画像間で共通する特徴を合意(consensus)として形成し，そのconsensusに最もマッチする領域を各画像の中から特定(localize)する．そして各画像でlocalizeされた領域の特徴をもとにして再度consensusを形成する，という一連の手続きを反復する．領域形状としては楕円を用い，同心楕円周上の輝度値に対するフーリエ変換によって特徴抽出を行う．また領域パラメータの探索にはMCMC法を用いる．実験により，比較的単純な画像群では対象の画像内領域をある程度推定することができ，提案手法の有効性が確認できた．

This paper discusses the importance of the norm selections for Computer Vision(CV) algorithms. L_p norm is often used as a dissimilarity measure between two vectors. For example, L_0 norm represents number of non-zero elements of a vector (assuming 0^0 ≡ 0), L_0 norm of a difference vector is called Manhattan Distance, and L_2 Norm is Euclidean distance. Because of the different poperties among these norms, CV algorithm can produce different result depending on the employed norm. We clarify that the following two problems are mainly caused by improper norm selections: "why Bag of Features based similar image search often finds the same image as similar to varieties of query images?" and "why discontinuous pixel values are estimated by discrete optimization based denoising algorithm employing smoothing energy term consisting of absolute sum of the neighboring pixel values?". Also, this paper shows these two problems are solved or relaxed just by changing the norms.

Anomaly detection can be applied to health monitoring of industrial plants, human medical conditions, vehicle conditions, and so on. A well-known anomaly detection method "Similarity Based Modeling" (SBM) proposed by Stephan W. Wegerich is acknowledged as an effective and essential method in this field. This report first shows that SBM is naturally derived as a special case of "Gaussian Processes" (GP) by regarding the similarity function in SBM as the kernel function in GP, where GP was proposed before the SBM patent. This fact provides us a new interpretation that GP is an example based nonlinear regression method. Based on this interpretation, next we proposed a method for relaxing the scalability problem of GP, that is, a method for reducing the size of gram matrix consisting of pairwise similarities on training samples. This method picks up the reference data for explaining the input data from the training dataset and dynamically construct "reduced" gram matrix. By using this matrix, we can estimate the output without losing the accuracy. The anomaly detection algorithm is applied to practical health monitoring problem of an electric plant with human operation and we confirmed that our algorithm successfully detect the pre-fault events before the actual fault.

This report proposes a fault and pre-fault detection method for industrial plants based on Gaussian process. Industrial plants can be monitored via attached sensors that measures temperature, pressure, voltage, electric current, and so on. Based on these sensor outputs, health monitoring of the target plant can be designed. The difficulty of this design problem is that the system fault can appear as statistical abnormality observed as irregular ensemble of the sensor outputs andlor temporal abnormality observed as irregularity of the time sequences. Furthermore, those systems are operated by human and it is difficult to distinguish the abnormalities caused by human-operation and system fault. We already proposed an abnormality detection system based on ICA and linear prediction, which avoids incorrect detections of abnormalities caused by human operations by recognizing and masking them. This method, however, cannot detect the system fault during the human operation. For solving this problem, we propose a unified method for statistical and temporal abnormality detection based on Gaussian Processes in this report. This method can detect system fault under normal human operation. We confirmed the effectiveness of our method through experiments on long-term sensory data sampled from real industrial plant.

ディスプレーを観測する人物の目の 3 次元位置が計測できれば，視点位置に依存した映像を提示することで，運動視差を作り出すことができ，3 次元物体の表示を行うことができる．我々は,能動ステレオカメラを用いて眼の位置を計測する方法を提案したが，この方法では装置が高価になってしまうため，一般に普及することは困難である．そこで，本研究では頭部に装着したカメラでディスプレーの矩形を観測し，この矩形の変形から実時間でカメラの位置と向きを計測する方法を提案する．この方法では，首振り機構を持たない単眼カメラのみでディスプレーを観測している際のカメラの位置・姿勢が計測できるため安価なシステムが構築可能である．ディスプレー枠の計測には位相計算型の勾配ベクトルを用いた Particle Filter を用いており，安定性と高速性の両者を同時に達成している．By using 3D position information of human eyes watching computer display, we can realize viewpoint dependent image displaying, i.e., motion parallax generation, which enables us popup displaying of 3D contents. We already developed a 3D displaying system using active stereo camera measuring 3D eye position. However, this system requires expensive equipments, such as, stereo camera, pan-tilt unit, rotary encoders, and so on. For making this 3D displaying method popular, we have to have a cheaper device that measures 3D eye positions. In this research, we use a head-mount camera observing LCD display frame. That is, we can estimate the head-mount camera position and orientation from the deformation of the display frame on the image. This method does not require pan-tilt unit, because the computer operator watches the display from any viewpoints and the head-mount camera captures the display frame. When the operator does not watch the display, we don't have to know the viewpoint position and orientation. In our implementation, we employed phase based gradient estimation for fast and robust implementation of particle filter..

本論文では，劣化画像を対象とした検索・照合の問題に適した画像間の相違性尺度を提案する．これまで，陰影の変化，遮蔽，ボケやブレ，解像度低下などの劣化に左右され難い検索・照合を実現するために，劣化の影響を受け難い特徴を使用することが研究されてきた．しかし，全ての劣化に対する不変特徴は存在せず，使用する特徴だけで劣化の影響を除去することはできない．本研究では，特徴ではなく相違性尺度を変更することによって，この問題を解決する方法を示す．劣化の多くは 「画像を直交展開した際の項の欠落」 によってモデル化することができる．このモデルの下では，欠落の起きていない部分の展開係数が一致すれば，劣化画像は比較した画像に近いと言える．この考えに基づき，まず，任意の直交展開に対して 2 枚の画像の展開係数の一致数ができるだけ多くなるように劣化画像の輝度値を調整する．次に，一致する要素が多いベクトル対に対してより小さい評価値が与えられる尺度を，上記の輝度調整された画像の展開係数ベクトルの組に適用する．このようにして得られた画像間相違度を用いた実験を通じて，提案する相違度により照合精度が，正規化相互相関を用いる場合よりも大幅に向上する事を確認した．This paper presents a set of dissimilarity measures suitable for image retrieval and matching tasks with degraded query images. For those applications, many works employ degradation invariant features. However, no feature can be invariant to all possible image degradations. In this paper, we focus on image dissimilarity measures rather than image features. A dissimilarity measure can be defined by answering two questions: "where to be measured between two image vectors" and "what metric to be used". Most image degradations can be modeled by partial masking of orthogonal image expansion coefficients. Under this model, the number of mismatching coefficient pairs can be a dissimilarity measure. Then, we first propose a method for adjusting coefficient vector magnitude so as to maximize the number of matching coefficient pairs. Next, by applying those measures proportional to the number of mismatching coefficients to the magnitude adjusted image pairs, we can define a degradation tolerant dissimilarity. Through extensive experiments, we confirmed that the matching rates by our measures are much higher than that of normalized correlation for image retrieval.

記事種別: 特集

近年では，特徴空間でのクラスタリングと画像空間での離散最適化を反復的に行い，各画素に最適なクラスタ中心値を割り当てる領域分割手法が提案されている．この方法は，特徴空間と画像空間の統合処理というだけでなく，新しい領域分割の枠組みへの見通しを示唆している．しかしながら，この方法は安定性と高速性に課題がある．この問題を解決するために，本研究では，特徴空間の分割と画像空間における離散最適化を行う方法を提案する．提案手法では，特徴空間の分割と Belief Propagation による最適化をいずれも階層的に行う．さまざまな画像を用いた実験では，精度と速度の観点で従来手法に対する有効性を確認することができた．Recently, a new image segmentation method has been proposed, that iteratively performs clustering in the feature space and discrete optimization in the image space, which assigns optimal cluster center values to pixels. This method provides us a vistas on novel image segmentation framework as well as a unified processing both in feature and image spaces. Unfortunately, this method has problems on stability and processing speed. For solving this problem, we propose a method that performs feature space partitioning and discrete optimization in the image space. The optimization by Belief Propagation and the feature space partitioning are done both in hierarchical manner. Through extensive experiments, we confirmed that our method performs better than existing method both in terms of accuracy and speed.

This report presents a set of dissimilarity measures suitable for image retrieval and matching tasks with degraded query images. Since local image features such as SIFT can easily be affected by global image degradations like motion blurring, we have to use pixel-by-pixel image dissimilarity measure. In this report, we confine the degradations to lacking of terms in arbitrary orthogonal expansion. In this case, if remained expansion coefficients match between two images, they should be the right pair. We first propose a method for computing sparse difference vectors consisting of coefficient differences for the right pairs. Next, a set of metric defined on vectors that take smaller values for sparser vectors while keeping the same L2 norm. Our dissimilarity measures are the combination of the difference vector computation and the sparsity sensitive metrics. Through the experiments, we confirmed that the one-by-one SQI matching rate by our measure is over two times bigger than that of normalized correlation for Extended Yale B Face database (subset 4).

In scanning duplex printed documents, show-through may degrade the image quality and character recognition rate. Show-through is a common phenomenon where the texts or images on one side is contaminated and appear on the other side. Since two images printed on both sides are essentially independent, this problem is a Blind Source Separation problem and one may think that Independent Component Analysis (ICA) can be applied for separating two images. However, ICA cannot produce show-through corrected images due to the frequent occurrence of the intensity pairs in the 2D space spanned by front side and backside intensities. In this report, we propose a show-through removal method using duplex scanned images by performing clustering and direct transformation of the 2D intensity distribution in order to make frontside and backside distribution independent. Through the separation experiments, we confirmed that our method performs well for removing show-through.

最近傍探索の研究においては，「次元の呪縛」 と呼ばれる現象が問題を困難にしてきた．この現象は，数学的に厳密に定義されていないが，「蓄積されるデータの分布の次元が一定の値を超えると，いかなるアルゴリズムでも全探索と等価になる」 という現象である．最近傍探索研究の歴史は，これを解決するために行われてきたと言っても過言ではない．これまでに数々の研究がなされてきたが，次元の呪縛が解けたという報告はこれまでになく，近年はこの現象を回避するための 「近似最近傍」 を探索する研究が盛んにおこなわれるようになってきた．本チュートリアルでは，近似を含むものと含まないもの両方について高速な最近傍探索アルゴリズムの解説を行う．その後に，果たして次元の呪縛を解くことが可能であるか否かについて再度検討を行う．The phenomenon so called "curse of dimensionality" makes the Nearest Neighbor (NN) search problem difficult and attractive. Without this phenomenon, NN search is just a boring problem. One-NN search problem is to find the closest pattern to a given query, and k-NN search is to find k-closest patterns. For solving these problems, so many accelerated algorithms have been proposed. However, the curse tells us that every accelerated NN search algorithm becomes linear (exhaustive) search when the stored data form a high dimensional distribution. Researchers who tackled this problem are sometimes regarded as daydreamers, because they fought with an unbeatable ghost. Recently, they changed their mind to stop this straight forward fighting. Instead, they are focusing on approximate NN search, which does not face the phenomenon. In this tutorial, some typical exact and approximate NN search algorithms are introduced, and we revisit the phenomenon, "curse of dimensionality" so as to solve this problem.

This paper presents a regression method, two-dimensional Mahalanobis distance minimization mapping (2D-M3), which is an extention of Mahalanobis Minimization Mapping: M3. M3 is a regression method between very high-dimensional input and output spaces based on Mahalanobis distance minimization criterion. Unlike the original M3, 2D-M3 directly extracts the features from image matrix rather than matrix-to-vector transformation. Because of this, 2D-M3 is much faster than original M3 without consuming much memory. We demonstrate the effectiveness of M3 through extensive experiments on a face image inpainting task.

Nearest Neighbor (NN) search plays important roles in example-based Computer vision algorithms. Accelerating NN search in very high-dimensional space, however, has a limitation, because of curse of dimensionality. For avoiding this problem, approximate NN search algorithms have been proposed. The most popular one is ANN which is basically a kd-tree based search algorithm with a feasible error. Recently, Locality Sensitive Hashing (LSH) is getting highlighted, because of its theoretical basis providing a clear relationship between the accuracy and the computational complexity. An improved hash based search algorithm, Principal Component Hashing (PCH), has been proposed, which is faster than ANN and LSH at the same accuracy without producing any search failures. However, most NN search algorithms share the limitations that 1) start-up time is mainly consumed for loading the data onto the memory, and 2) NN search for huge data cannot be executed, because the main memory size is limited. For solving this problem, we propose an external NN search algorithm rithm which directly finds the approximate NN data on HDD based on PCH algorithm. The basic idea is simple. By replacing the hash bins on a projection axis by a file having B+ tree structure, we can realize memory efficient PCH which works on HDD. In the experiment, we confirmed that the algorithm can perform NN search on huge database, which cannot be loaded on main memory. Also, we noticed that the search algorithm gains unexpected acceleration by the cashing mechanism of Linux operating system that frequently accessed data on HDD is kept on the memory.

This paper brings out a robust pixel-wise object tracking algorithm which is based on the K-means clustering algorithm. In order to achieve the robust object tracking under complex condition (such as wired objects cluttered background) a new reliability-based K-means clustering algorithm is applied to remove the noise background pixel (which is neigher similar to the target nor the background samples) from the target object. According to the triangular relationship among an unknown pixle and its two nearest cluster centers (target and background) the normal pixel (target or background one) will be assigned with high reliability value and correctly classified while noise pixels will be given low reliability value and ignored. A radial sampling method is also brought out for improving both the processing speed and the robustness of this algorithm. According to the proposed algorithm we have set up a real video-rate object tracking system. Through the extensive experiments the effectiveness and advantages of this reliability-based K-means tracking algorithm are confirmed.This paper brings out a robust pixel-wise object tracking algorithm which is based on the K-means clustering algorithm. In order to achieve the robust object tracking under complex condition (such as wired objects, cluttered background), a new reliability-based K-means clustering algorithm is applied to remove the noise background pixel (which is neigher similar to the target nor the background samples) from the target object. According to the triangular relationship among an unknown pixle and its two nearest cluster centers (target and background), the normal pixel (target or background one) will be assigned with high reliability value and correctly classified, while noise pixels will be given low reliability value and ignored. A radial sampling method is also brought out for improving both the processing speed and the robustness of this algorithm. According to the proposed algorithm, we have set up a real video-rate object tracking system. Through the extensive experiments, the effectiveness and advantages of this reliability-based K-means tracking algorithm are confirmed.

In thie paper, a mirrored SIFT feature is proposed for detecting plane-symmtric objects. This feature can be used to evaluate the symmetric-pair likeness of feature points of a pseudo affine transformed plane-symmetric object. Plane-symmetric object detection is carried out by using the proposed mirrored SIFT feature and scale feature. The effectiveness has been confirmed through several experiments using real images.

This paper presents a method for detecting and tracking unknown number of targets in image sequences based on EM algorithm. EM algorithm can fit mixture of distributions to observed data. Applying this algorithm to the foreground pixels obtained by background subtraction, we can estimate the assumed number of object positions, and shape parameters. Also, we can estimate the number of objects based on MDL criterion, where the valid number of objects provides the minimum DL. A simple combination of EM algorithm and MDL criterion, however, will produce incorrect object number. This is because the constraint on shape and size of each object is too loose and EM algorithm may estimate inconsistent fittings. We propose a robust method for counting the number of objects by introducing a constraint between the object position and shape parameters, i.e., covariance matrix. This constraint can be represented by a hyper plane in 5-D space spanned by positions(x,y) and covariance parameters(σ^2_x,σ_xy,σ^2_y). By confining the distribution parameters on this plane in EM algorithm, we can robustly estimate the number of objects.

The Nearest Neighbor(NN) search can be applied to the fixed-dimensional vector data with a metric satisfying axioms of distance. On the other hand, maximum-similarity search can be applied to wide variety of data and metrics including fingerprint matching, where the data cannot be represented as fixed-dimensional vectors and the similarity measure cannot be converted to distance measure. Many methods have been proposed for solving this problem. Among them, Maeda's method has been well known as one of the fastest search algorithm. The method, however, uses matching score matrix representing similarity measures of all possible combinations of n stored data, which consumes O(n^2) memory. Hence, it is almost impossible to perform maximum similarity search on millions of data. In this report, we propose a method for extracting independent data among stored data. Stored data and the query can be mapped to a fixed dimensional space by using the matching scores with the extracted data. Then we apply our NN search algorithm Principal Componen Hashing(PCH) for performing approximate maximum similarity search. Through some experiments, we prove that our algorithm outperforms on accuracy and speed.

We present a method for adequate alignment of non-target cluster centers for K-means tracker algorithm. In this method, the number of non-target cluster centers and alignment of them are determined based on the distances from a pixel which is on the ellipse surrounding a target to both target cluster centers and non-target cluster centers. This distance is calculated in the 5 dimensional feature space used in K-means tracker. This method can prevent selecting the pixel as a non-target cluster center which doesn't contribute to clustering for tracking a target or which has similar feature with other determined non-target cluster centers. Through a comparative object tracking experiment, we confirmed that this algorithm contributes stabilization of object tracking and improvement of efficient processing.

Recently, AdaBoost is a general algorithm for the face detection. Since AdaBoost cannot generate new features, it is necessary to prepare a large amount of features beforehand. The detection performance depends on the prepared features. In this paper, new features are generated by GA, and the efficient features are selected by evaluating both independence between features and conviction calculated with Real AdaBoost. Then Real AdaBoost constructs the classifier by using those features.

This paper brings out a robust pixel-wise object tracking algorithm which is based on the K-means clustering. In this paper, the target object is assumed to be non-rigid and may contain apertures. In order to achieve the robust object tracking against such objects, several ideas are applied in this work: 1) Pixel-wise clustering algorithm is applied for tracking the non-rigid object and removing the mixed background pixels from the search area; 2) Embedding the negative samples into K-means clustering so as to achieve the adaptive pixel classification without the fixed threshold; 3) Representing the image feature with a color-position feature vector so that this algorithm can follow the changes of target colors and position simultaneously; 4) A variable ellipse model is used to restrict the search area and represent the surrounding background samples; 5) Tracking failure detection and recovery processes are brought out according to both the target and background samples; 6) A radial sampling method is brought out not only for speeding up the clustering process but also improving the robustness of this algorithm. We have set up a video-rate object tracking system with the proposed algorithm. Through extensive experiments, the effectiveness and advantages of this K-means clustering based tracking algorithm are confirmed.

We propose a new self-localization method for a mobile robot equipped with a fixed-viewpoint active camera by observing multiple parallel lines patterns with the camera. The method utilizes the technique for estimating location and posture of the camera from one parallel lines pattern which was previously developed by our research group. The method switches from a certain parallel lines pattern used for self-localization to another parallel lines pattern while the robot is moving in the environment. After the new parallel lines pattern is selected, the method keeps estimating location and p...

This paper reviews and surveys example based techniques in the fields of Pattern Recognition and Computer Vision, from the viewpoints of anormality detection, classification, non-linear mapping learning, regularization, and basics: nearest neighbor search and non-parametric probability density estimation.

This paper reviews and surveys example based techniques in the fields of Pattern Recognition and Computer Vision, from the viewpoints of anormality detection, classification, non-linear mapping learning, regularization, and basics: nearest neighbor search and non-parametric probability density estimation.

This report gives an overview of the 10th IEEE international conference on computer vision (ICCV2005), which was held in Beijing, China, from October 17th to 20th, 2005.

This paper presents a novel method for human head tracking using multiple cameras. Most existing methods estimate 3D target, position according to 2D tracking results at different viewpoints. This framework can be easily affected by the inconsistent tracking results on 2D images, which leads 3D tracking failure. For solving this problem, we investigate a natural extension of CONDENSATION to multi-viewpoint images. Our method generates many hypotheses on a target (human head) in 3D space and estimates the likelihood of each hypothesis by integrating viewpoint dependent likelihood values of 2D hypotheses projected on image planes. In theory, view point dependent likelihood values should be integrated by multiplication, however, it is easily affected by occulusions. Thus we investigate this problem and propose a novel integration method in this paper and implemented a prototype system consisting of six set of PCs and cameras. We confirmed the robustness against occulusions and the efficiency of our method.

In laser arc hybrid welding of a galvanized steel sheet, I examined influence of the shielding gas which generated a welding defect. By welding phenomena observation, it was confirmed that a defect occurred by a difference of molten pools behavior.

This paper presents a new framework for learning dynamic robot behavior such that a wheeled mobile robot moves to the destination with wheel slippage. First, our method learns the body dynamics from a large number of sensory-motor instances using PaLM-tree algorithm, which approximates nonlinear mapping in arbitrary preciseness. The mapping specifies a transition model from state-action pair to next state. An optimal action sequence generating the transition from initial to goal states is obtained by the branch-and-bound algorithm using the estimated transition model. Then, new sensory-moto...

This paper presents a K-means tracker, which is a novel visual tracking algorithm. This algorithm is robust against the interfused background, because it discriminates "target" pixels from "background" pixels in the tracking region by applying K-means clustering to both the positive and negative information of the target. To ensure the robustness of this algorithm, we apply the following ideas : 1) To represent the color similarity and spatial approximation of the target simultaneously, we use a 5D feature vector consisting of the position (x, y) and color (y, u, v) information for object tracking. The object tracking is performed and updated not only in the image space but also in the color space at the same time. Therefore, this adaptive nature guarantees our method to be robust against the target color change. 2) By using a variable ellipse model to restrict the target search area and represent the non-target pixels surrounding the target, the algorithm can cope with the changes of the scale and shape of the target object flexibly. 3) Even the tracking sometimes fails, this algorithm can automatically discover and recover from the tracking failure based on the positive and negative information. To capture motion-blur-free images of a moving object at video rate, we control a set of active cameras which are mounted on the pan-tilt units according to the results of the K-means tracker.

We have proposed an algorithm for estimating the visual direction using iris contours, which are detected from an input image and fitted with ellipses. However, it is difficult to detect the iris contours without the boundaries between an eyelid and the iris. In order to solve this problem, we propose an simple eye-model that consists of the iris contours and the eyelids. This paper describes how to track the iris contours correctly and robustly by using the eye-model. We use Condensation algorithm for tracking stably the iris contours against blinking and propose a likelihood function based on directions of iris contours and a brightness of iris region.

Fractal coding is based on the assumption of self-affinity. In image coding, blocks of an image can be approximated by extracting larger blocks from elsewhere in the image. For that reason, in the plane with a homogeneous texture pattern, a front image can code more efficiently than the inclined image. In this paper, we examine the index of coding efficiency when performing image compression with fractal picture coding. And then, we propose an algorithm to estimate the plane orientation. Through various experiments with generated images and real images, we confirmed our method is effective.

This paper presents evolutional history of regression trees, a novel regression tree called PaLM-tree, and its application to Computer Vision and Pattern Recognition problems. Regression tree is a tree representation of nonlinear function, which is mainly investigated in the fields of Machine Learning and Data Mining. The original regression tree is a simple tree representation holding function values at leaves (terminated nodes). Linear regression tree extends the function values to linear regression coefficients. Since these regression trees are designed for rule-extraction, most of them ...

In this paper, we propose a new method called structuralized MCMC, which is an extension of the framework of MarRov Chain Monte Carlo (MCMC). Our method estimates parameters of the objects that consist of some parts whick have structural relationship among themselves. Parameters of objects are divided for each part, and the structure among parts is represented as a conditional probability on distributions of other parts. The MCMC is applied by using the conditional probability as a kernel of MarRov-chain. Each part is estimated by using the estimated distribution of other parts. Our method can estimate the distribution of each part while considering the structure among parts. This Structuralized MCMC is applied to a facial parts detection problem. The extensive experiments demonstrate the effectiveness of our method.

This paper presents a perspective imaging model that can be used to represent any single viewpoint imaging system. Our imaging model consists of an optical center, whose position is well determined in the 3D space, and a virtual screen on which undistorted images taken with the uncalibrated camera are displayed. This model uses two screens as virtual image planes, and the optical center is estimated as the converging point of those rays that are passing through the corresponding points on these planes. This method has the advantages that 1) it can be applied to any single viewpoint camera, and 2) it can remove any type of distortions. In the experiments, results for normal and omni-directional cameras demonstrate the effectiveness of our method.

In this paper, we describe a novel method that can estimate the extrinsic parameters and the focal length of a camera by using only one single image of two coplanar circles with arbitrary radius. This method does not require that the circle centers or the full circles are viewable, and the circles can have any unknown radius. Compared with conventional methods, our method has the following advantages : 1) it does not use specially designed objects, 2) it can obtain unique solution by using only one monocular image, 3) it can estimate the focal length and the extrinsic parameters, and 4) it can be used to estimate the relative translation and rotation between cameras. The extensive experiments over simulated images and real images demonstrate the robustness and the effectiveness of our method.

This report presents an accelerated Nearest Neighbor (NN) classifier for quadtree representations of binary images generated by color target detection. This method finds the NN prototype to the input quadtree from many prototypes and classifies the input to the prototype class. For this purpose, we add the density information to the grey nodes of the tree. In the coarse-to-fine comparison between two trees, we can calculate the upper and lower bounds of the distance between these trees by referring the density information at any level. Using these upper and lower bounds, we can reduce the NN candidates to the input in the similar way to the branch-and-bound or A* algorithms. We modified this method by performing best-first search for accelerating the decreasing speed of the minimum upper bound. This enables further reduction of the NN candidates. The classification speed can further be accelerated depending on the fact that the problem is not a NN search problem but a NN classification. That is, if all the NN candidates belong to a single class, we can classify the input immediately. Through the experiments, we confirmed that our method is over forty times faster than the brute force NN search.

Recently mass-mail sending is becoming a serious problem that prevents normal commerce and advertisement on the Internet. We are now developing a system solving this problem. Existing mass-mail filtering systems can be classified into two types; 1) open relay MTA based filtering systems and 2) recognition based contents filtering systems. Since the number of open relay MTA is much bigger than the number of mass-mail contents, contents based mass-mail filtering is much effective than the open relay based filtering. However, the recognition based contents filtering requires time consuming training and tuning. Unlike these approaches, our system consists of two parts; mass-mail checksum database and clients. The checksums are delivered via the DNS and the clients. The prototype database system is currently running, i.e., mass-mail contents sent to some dummy addresses are analyzed, normalized, and their checksums are stored into the DNS. As well, some clients referring the DNS are developed. This report presents the overview of our current system and the future extensions.

Recently mass-mail sending is becoming a serious problem that prevents normal commerce and advertisement on the Internet. We are now developing a system solving this problem. Existing mass-mail filtering systems can be classified into two types; 1) open relay MTA based filtering systems and 2) recognition based contents filtering systems. Since the number of open relay MTA is much bigger than the number of mass-mail contents, contents based mass-mail filtering is much effective than the open relay based filtering. However, the recognition based contents filtering requires time consuming training and tuning. Unlike these approaches, our system consists of two parts; mass-mail checksum database and clients. The checksums are delivered via the DNS and the clients. The prototype database system is currently running, i.e., mass-mail contents sent to some dummy addresses are analyzed, normalized, and their checksums are stored into the DNS. As well, some clients referring the DNS are developed. This report presents the overview of our current system and the future extensions.

For the camera calibration, point correspondence data between 3D and 2D points widely spreading over the image plane is necessary to get a precise result. However, in the case of wide observing area, it is impractical to use a huge calibration object that covers the camera view. This paper presents a refined calibration method using circular patterns. We already proposed a circular pattern based camera calibration method using a four-wheel mobile robot. That is, by keeping low speed and constant steering angle, the robot can draw a big circular pattern on a plane. Our previous method, however, has a limitation that roll angle around the optical axis is assumed to be zero. This paper presents an improved version of the camera calibration method. Our method estimates the extrinsic parameters of fixed camera from a single circular trajectory when the focal length of the camera is known. In the case that the focal length is not given, two co-planar un-concentric circular trajectories are enough for calibrating the extrinsic parameters and the focal length. In both cases, the center and the radius of the circle(s), the speed of the robot are not required. The extensive experiments over simulated images as well as real images demonstrate the robustness and the accuracy of our method.

高精度のカメラキャリブレーションを実現するために、3次元位置が既知である校正用の基準点はできるだけ画像面に広く分布した方が望まれる。しかし、カメラの観測範囲は運動場のように広大である場合、全画面がカバーできる巨大な物体を用意することが非常に困難である。我々は、4輪車輌型ロボットを用いて、真円の軌跡を作り、それを利用したカメラキャリブレーション法を既に提案した。しかし、その方法にはカメラ光軸回りの回転がないという制限があった。本稿では、円パターンを利用した完全なカメラキャリブレーション法を提案する。本方法では、円心と半径、そしてロボットの移動速度の情報を使わずに、１）焦点距離既知の場合、一個の円パターンからカメラの外部パラメータを推定する；２）焦点距離未知の場合、同じ平面内における二つの非同心円のパターンからカメラの外部パラメータと焦点距離を求める方法を提案する。シミュレーションと実画像を用いた実験により提案手法の有効性を確認した。For the camera calibration, point correspondence data between 3D and 2D points widely spreading over the image plane is necessary to get a precise result. However, in the case of wide observing area, it is impractical to use a huge calibration object that covers the camera view. This paper presents a refined calibration method using circular patterns. We already proposed a circular pattern based camera calibration method using a four-wheel mobile robot. That is, by keeping low speed and constant steering angle, the robot can draw a big circular pattern on a plane. Our previous method, however, has a limitation that roll angle around the optical axis is assumed to be zero. This paper presents an improved version of the camera calibration method. Our method estimates the extrinsic parameters of fixed camera from a single circular trajectory when the focal length of the camera is known. In the case that the focal length is not given, two co-planar un-concentric circular trajectories are enough for calibrating the extrinsic parameters and the focal length. In both cases, the center and the radius of the circle(s), the speed of the robot are not required. The extensive experiments over simulated images as well as real images demonstrate the robustness and the accuracy of our method.

This paper presents two efficient condensing algorithms for nearest neighbor classifiers. Condensing is to remove prototype patterns keeping the classification accuracy of nearest neighbor classifiers. This method plays important roles not only for the nearest neighbor classifiers, but also for the support vector machines, because the resulting prototype patterns involve support vectors in many cases. However, the previous condensing algorithms are computationally inefficient in general Pattern Recognition tasks. This is because it refers proximity graphs, e.g., Delaunay and Gabriel graphs, which consume exponential computational time for the dimension of pattern space. We propose two condensing methods that do not require the entire proximity graph of given patterns.

We have proposed new theorems for efficient condensing algorithms based on proximity graphs. This paper presents implementations and evaluations of two condensing algorithms, called Direct condensing and Chip-off condensing, based on our theorems. The algorithms can efficiently obtain the same prototypes as the Voronoi condensing without the entire proximity graph of given patterns. The Direct condensing algorithm is more computationally efficient than any previous condensing algorithms for voronoi condensing, and the Chip-Off condensing algorithm can sequencially remove disused patterns while keeping the classifier boundaries. This paper shows the evaluations of proposed condensing algorithms, and demonstrates applications of the nearest-neighbor classifier and the condensing.

The color competition method is proposed as a method of detecting the specified color from an image. This is the method of constituting a look up table using nearest neighbor in a color space. This method is known to be able to learn and classify fast and accurately if the method be trained by specifying target and nontarget color examples. This method can accurately detect if there are many training patterns. However, it is difficult to collect many training patterns manually. In this report, we propose a method that collect training patterns of skin color and non-skin color, tracking humans based on skin color, and learn skin color dynamically. This method generates hypotheses of training patterns based on regions of tracking faces and continuity in color variation, and verifies these hypotheses based on similarity of colors and contribution to detecting skin color, and learn likely training patterns.

This paper presents a novel camera calibration method for RC car robot control system with external cameras. This system can be regarded as a sensorless robot system, which has advantages of low-cost body and precise localization by external camera. The only drawback is that it requires a priory camera calibration. For this calibration, we don't use any specially designed object, but the robot is employed as a calibration object. Before the calibration, the robot cannot move along a straight line, however, it can move along a circle by keeping constant steering angle and slow speed. The circular locus of the robot is projected as an ellipse on the image plane by the perspective projection, Based on this geometric relationship, we propose a calibration method of the focal length and the view direction of the camera relative to the working plane of the robot. Through computer simulations and experiments with a real camera and a robot, we have confirmed that our method is more robust than the Homography based calibration using four point correspondences.

This paper presents a method for map generation based on the interaction between a robot body and its surrounding environment. While a robot moves in the environment, the robot interacts with its surrounding environment. If the effect of the environment on the robot changes, such interactions also change. By observing the robot's body, our method detects such change of the interaction and generates a description representing the type of change and the location where such change is observed. In the current implementation, we assume that there are two types of the change in the interaction. The real robot experiments are conducted in order to show the validity of our method.

Most nearest neighbor (NN) classifiers employ NN search algorithms for the acceleration. However, NN classification does not always require the NN search. Based on this idea, we propose a novel algorithm named k-d decision tree (KDDT). Since KDDT uses Voronoi condensed prototypes, it is less memory consuming than naive NN classifiers. We have confirmed that KDDT is much faster than NN search based classifiers through the comparative, experiment (from 9 to 369 times faster).

This paper presents a novel background subtraction method for detecting foreground objects in dynamic scenes involving swaying trees and fluttering flags. Most methods proposed so far adjust the permissible range of the background image variations according to the training samples of background images. Thus, the detection sensitivity decreases at those pixels having wide permissible ranges. If we can narrow the ranges by analyzing input images, the detection sensitivity can be improved. For this narrowing, we employ the property that image variations at neighboring image blocks have strong correlation, also known as "cooccurrence". This approach is essentially different from chronological background image updating or morphological postprocessing. Experimental results for real images demonstrate the effectiveness of our method.

Color target detection tasks can ha regarded as a color classification problem. Most of the classification methods first integrate features into a single value representing the probability or similarity, and the classification is done by comparing or thresholding these values. This integration, i.e. mapping from multi-dimensional pattern space to 1-dimensional space removes the manifold of pattern distributions in the original space, and hence, the exceptional patterns can be missclassified. In this report, we propose a novel method for color target detection based on nearest neighbor classifier, which doesn't require the feature integration and can be trained by specifying target and nontarget color examples. Since the color classifier is implemented as LUT, our system can detect targets in video-rate. Experimental results demonstrate the effectiveness of our method.

In this panel, we will discuss varieties of topics, e.g., "What is the intrinsic information obtained only by mass-camera system?", "What is the role of 'communication' in mass-camera system?", "Future systems based on mass-camera". Through the discussions, we hope to have a vision of the future in this research field.

A rotational camera can be regarded as a sensor that accumulates and integrates multiple image information acquired by changing the camera parameters, i.e., pan, tilt, zoom, exposure, and so on. The accumulation and integration should be performed for each view direction from a single view point. By using a rotational camera, we can realize 1)precise viewpoint calibration, 2)high spatial resolution, 3)wide dynamic range of pixels, and 4)high S/N ratio. As for the omnidirectional imaging, a rotational camera must rotate to observemultiple view directions, i.e., an omnidirectional image cannot be captured at once. However, as an active camera, its time response is equivalent to that of an ordinary video camera. Especially, a specially calibrated rotational camera whose projection center coinsides with its rotation center has advantages that 1)background subtraction can be performed for the captured images by a rotating camera, 2)simple egomotion analysis can be applied, because camera action will not cause motion parallax. In this paper, all above topics are discussed.

本論文では, 3次元形状復元の実用的問題の一つとして, イメージスキャナによって撮影された書籍表面の陰影画像から, 固有空間法を用いて高速に書籍の断面形状を復元し, 画像中の陰影とひずみを補正する方法について述べる.具体的には, 互いに対をなす陰影画像と断面形状のデータを事前に多数取得し, これらを正規化して別々の固有空間に保存する.そして, 与えられた画像に対して陰影固有空間内で局所的な線形補間を行い, この係数を用いて形状固有空間から断面形状を復元し, 画像の補正を行う.ここでは, 固有空間中の局所的な線形補間での誤差と, 陰影固有空間から形状固有空間への線形変換での誤差を考慮し, データに対するSN比の概念を導入して線形補間における係数を求める.また, データの正規化として, フーリエ変換を用いて書籍の平行移動成分を除去する方法を用いる.提案した手法の妥当性を示すために, 模型及び実際の書籍を用いた実験を行い, 高速で精度の良い形状復元と画像の補正が可能であることを確認した.

We propose a method for detecting and tracking a human head in real time from an image sequence. The proposed method has three advantages: (1) we employ a fixed-viewpoint pan-tilt-zoom camera to acquire image sequences
with the camera, we eliminate the variations in the head appearance due to camera rotations with respect to the viewpoint
(2) we prepare a variety of contour models of the head appearances and relate them to the camera parameters
this allows us to adaptively select the model to deal with the variations in the head appearance due to human activities
(3) we use the model parameters obtained by detecting the head in the previous image to estimate those to be fitted in the current image
this estimation facilitates computational time for the head detection. Accordingly, the accuracy of the detection and required computational time are both improved and, at the same time, the robust head detection and tracking are realized in almost real time. Experimental results in the real situation show the effectiveness of our method. © 2000 IEEE.

Silhouette volume intersection is one of the most popular ideas for reconstructing the 3D volume of an object from multi-viewpoint silhouette images. This paper presents a novel parallel volume intersection method based on plane-to-plane homography for real-time 3D volume reconstruction using active cameras. This paper mainly focuses on the acceleration of back-projection from silhouette images to 3D space without using any sophisticated software technique, such as octree volume representation, or look-up table based projection acceleration. Also this paper presents a parallel intersection method of projected silhouette images. From the preliminary experimental results we estimate near frame-rate volume reconstruction for a life-sized mannequin can be achieved at 3cm spatial resolution on our PC cluster system.

This paper addresses active image capturing and dynamic scene visualization by Cooperative Distributed Vision (CDV, in short). The concept of CDV was proposed by our five years project starting from 1996. From a practical point of view, the goal of CDV is summarized as follows: Embed in the real world a group of network-connected Observation Stations (real time video image processor with active camera(s)) and mobile robots with vision. And realize 1) wide-area dynamic scene understanding and 2) versatile scene visualization. Applications of CDV include real time wide-area surveillance, remote conference and lecturing systems, interactive 3D TV and intelligent TV studio, navigation of (non-intelligent) mobile robots and disabled people, cooperative mobile robots, and so on. In this paper, we first define the framework of CDV and give a brief retrospective view of the computer vision research to show the background of CDV. Then we present technical research results so far obtained: 1) fixed viewpoint pan-tilt-zoom camera for wide-area active imaging, 2) moving object detection and tracking for reactive image acquisition, 3) multi-viewpoints object imaging by cooperative observation stations, and 4) scenario-based cooperative camera-work planning for dynamic scene visualization. Prototype systems demonstrate the effectiveness and practical utilities of the proposed methods.

For the detection and tracking of objects moving in wide surveillance area, active sensing method using pan-tilt-zoom camera is effective. In the visual reactive system which autonomously changes camera barameters depending on the visual information, time elapses while capturing image, determining and performing camera action. Because of this elapsed time, detected attributes (position, speed, etc.) of objects in the system correspond to the past attributes in the real world. This time lag causes inconsistent camera action with moving objects. Hence, the realtime system should perform consistent camera action with moving objects by canceling this time lag. This report presents a method to determine the consistent camera action based on 1) motion estimation of object and 2) response-time property of the system. Experimental results demonstrate that the proposed method drastically improves the performance and stability of the object tracking system using fixed viewpoint camera.

Most of behavior recognition methods proposed so far share the limitations of bottom-up analysis, and single-object assumption; the bottom-lip analysis can be confused by erroneous and missing image features and the single-object assumption prevents us from analyzing image sequences including multiple moving objects. This paper presents a robust behavior recognition method free from these limitations. Our method is best characterized by I) top-down image feature extraction by selective attention mechanism, 2) object discrimination by colored-token propagation, and 3) integration of multi-viewpoint images. Extensive experiments of human behavior recognition in real world environments demonstrate the soundness and robustness of our method.

We address the problem of recovering the 3D shape of an unfolded book surface from the shading information in a scanner image. This shape-from-shading problem in a real world environment is made difficult by a proximal, moving light source, interreflections, specular reflections, and a nonuniform albedo distribution. Taking all these factors into account, we formulate the problem as an iterative, non-linear optimization problem. Piecewise polynomial models of the 3D shape and albedo distribution are introduced to efficiently and stably compute the shape in practice. Finally, we propose a method to restore the distorted scanner image based on the reconstructed 3D shape. The image restoration experiments for real book surfaces demonstrate that much of the geometric and photometric distortions are removed by our method.

This paper presents a motion-recognition method for visual-surveillance tasks in Cooperative-Distributed-Vision(CDV) system. This method is designed as a motion recognition mechanism of Observation Station, so as to be capable of the real-time recognition of multiple-object motions. To realize multiple-motion recognition, we propose Selective Attention Model, Which iterates event detection in focusing regions on the image space and the action renewing the focusing regions according to the event detection result. While iterating the event detection and action, the model changes its internal states corresponding to the object-motion stage. By employing an NFA as its state transition mechanism, the model can generate and follow all possible states. The action of this model can be expanded as "camera-parameter control" and "message sending to other stations", which enables active and cooperative motion recognition.

Research field for PRMU researchers is expanding rapidly. Recent criticisms to mathematical model researches for compter vision was summarized as follows: (1) Can we find new directions for mathematical model? Is it exciting? (2) Does it work in real world? Is it useful for real applications? In the panel discussion we will review the above criticisms and discuss what comes next for mathematical model researches.

Background subtraction is a simple and effective method to detect anomalous regions in images. In spite of the effectiveness, it cannot be used with an active (moving) camera head because the background image varies with camera-parameter control. This paper presents a background subtraction method with pan-tilt-zoom control. The proposed method consists of an omnidirectional background model called appearance sphere and parallax free sensing. Based on this model, precise background images can be generated and background subtraction can be performed for any combination of pan-tilt-zoom parameters without restoring 3D scene information. © 1996 IEEE.

This report describes an overview of the fifth International Conference on Computer Vision, which was held at MIT, USA in June 20-23, 1995.

第5回目を迎えたInternational Conference on Computer Vision1995(ICCD'95)は, 今回は6月20日〜23日の間, 米国マサチューセッツ工科大学で開催された.会議の主催者は, IEEE, PAMI技術委員会が母体となって組織されたICCV'95実行委員会である.会議の規模は, 応募論文数600件, 採択論文数161件(口頭56体, ポスタ105件), 会議参加者582人であった.ICCVは, 2年に1回の間隔で開催されるコンピュータビジョンの国際会議であり, 会議の規模自体は比較的小規模であるが.論文審査は非常に厳しく, セッションはシングルトラックで行われるため, 発表や議論の質に関しては研究者間でも定評がある.

γ-ω Hough transform, which we proposed before, has the following advantages over the ordinary ρ-θ Hough transform: (a)The ρ-θ parameter space involves the inherent bias, i.e. the uneven sensitivity to directions of straight lines. In the γ-ω space this bias is completely eliminated. (b)A voting curve in the γ-ω space is a piece-wise linear line consisting of two segments. This enables the fast voting as well as facilitates the analysis of geometric properties of straight lines. In practice, however, the γ-ω Hough transformation algorithm involves a crucial problem: votes from those pixels on the same digital line are not always accumulated into a single cell, quantized unit, in the parameter space and hence spread over several neighboring cells. The essential cause of this problem is that the set of cells in the γ-ω space does not represent all possible digital lines in an image. In this paper, we first analyze detailed geometric properties of digital lines and propose a new quantization method, i.e. cell configuration, of the γ-ω space. In this cell configuration, all possible digital lines have one-to-one correspondence to the set of cells. Then, considering the uncertainty range of vote from a pixel, a new voting method is proposed. We call the Hough transformation algorithm incorporating these new cell configuration and voting method "high precision γ-ω Hough transformation algorithm". Experimental results have demonstrated that digital lines of arbitrary orientations and location can be correctly detected by the high precision γ-ω Hough transformation algorithm.

本論文では、イメージ・スキャナによって書籍表面を撮影した画像から、書籍表面の3次元形状を復元する問題を対象とし、（）照明光の強度、および光源からの光の方向がスキャナ面から書籍表面までの距離に応じて変化する（）書籍表面の反射特性が拡散反射と正反射を含むPhongの反射モデルに従うという2点を考慮した解析モデルを構成することによって、精度良く3次元形状が復元できることを示す。推定された3次元形状を用いることにより、画像中の陰影、幾何学的歪みを補正することができる。実験の結果、スキャナ画像から真の3次元形状にほぼ一致した形状が復元され、補正後の画像の可読性は原画像と比べて大幅に改善されていることが確認され、本手法の有効性が示された。In this paper, we present a method to estimate the 3D shape of an unfolded curved book surface from an image taken by an image-scanner. To solve this estimation problem, we augmented the ordinary Shape from Shading model in the following two points. The first augmentation is that the illuminant intensity and the lighting direction vary with the gap between the scanning plane and the book surface. The second augmentation is that the Phong's reflection model is incorporated to cope with the specular reflection on the book surface. In the experiments, the 3D shape of the book surface was correctly reconstructed using the augmented model. Then the estimated 3D shape was used for the restoration of the distorted scanner image. This image restoration considerably improved the readability of the book surface.

本稿では,オンラインで利用可能な多数のデータベースを対象とした,データベース検索支援システムについて述べる。近年、データベースの普及に伴ない,データベースの種類およびデータベース数が増加している。そのため、データベース群全体を把握し,検索要求に応じる適切なデータベースを選択することが因難になっている。我々は,データベース群について十分な知識を有さないユーザが,検索要求に適したデータベースを選択するために必要となるメタ情報および基本機能を示した。メタ情報は,各データベースの特徴および実際の内容に関する情報であり,メタ情報に対して基本機能を適用させることにより,適切なデータベースが選択される。本稿では,メタ情報の分類のための指針を示し,さらに,メタ情報の生成,および,利用方法について述べる。

波形間の対応する部分を求めるという波形の照合問題は,信号処理やステレオ・ビジョンなどの分野における重要な課題である.波形の凹凸領域は,尺度空間フィルタリングを用いて階層的にとらえることができる.このような階層構造を対応付けることにより,安定で効率のよい対応付けが可能になる.そこで本報告では,波形の階層構造そのものを対応付ける「構造照合」という問題を提案し,2次零交差線の包含関係を用いて波形の階層構造を表わした場合の,構造照合の定式化と解法について論じた.

2つの波形間の対応を求めるという波形の照合問題は，信号処理をはじめとする様々な分野における重要な課題の1つである．従来この問題に対しては，波形上の特徴点を直接対応付けるための手法が，主に検討されてきた．波形の凹凸構造は本来，階層構造を持っている．この階層構造を対応付けることにより，精度の良い，しかも安定した対応付けが可能になると考えられる．本報告では，波形の階層構造に着目し，この階層構造を対応付ける「構造照合」という問題を新たに提案する．まず，尺度空間フィルタリングにより得られる2次零交差線の，包含関係を用いて波形の階層構造を表わす．そして，この階層構造を構造点の木構造として記述し，波形間の形態の変動構造を確率論的に表現することで，波形の構造照合を最尤推定問題として定式化する．さらに，この問題の解法を論じる．The pattern matching of waveforms is an important problem in the fields of signal proccesing, image processing, and so no. Hitherto some methods have been proposed, which match the feature points of waveforms directly. Since the convex and the concave regions of waveform have hierarchical structure, it is necessary for a stable and precise matching to match the hierarchical structure. In this paper, we propose a structural matching problem that matches the hierarchical structure of scale-space filtered waveforms, and formalize this problem as a likeliwood estimation of fluctuation of structure points.

本研究では数本の木が配置された二次元環境中を移動するカメラによって環境を観測した場合の画像を想定し，スクリーン上の一次元濃淡波形を「一般化波形の構造線」を用いて解析した．構造線は波形の凹凸領域の階層構造をあらわす解析的な曲線であり，位相的には三分木の構造を持っている，解析の結果，構造線の枝の生成・消滅，切断・接合などの形態変化と木の見えかくれ，カメラの視界への対象の出入り，スクリーン上での木の位置関係，大小関係の変化が関係あることがわかった．この関係を整理することによって観測された画像中での木の見えかくれなどのシーンの見え方の変化を判断することが可能である．Structure line is a hierarchical representation of waveforms based on scale space filtering. Structure line has the same topological property as the trinary tree, and represents the hierarchy of the convex and concave region of the waveform. In this paper, we discuss an application of the structure line to the computer vision. One of the most basic and difficult problem of the computer vision is the reconstruction of the 3-dimensional object from the multi-viewpoint images. There may be the occlusion of the characteristic points between the images, then we can't find the matching pair. So, we shold know the occluded regions between the images before the beginning the correspondence process. We investigated the relation between the morphological transition of the structure line of the observed images and the transition of the scene. The relation between the transitions of the structure line and the scene transition are cleared.

type:text

小規模化の際におけるニューラルネットワークの性能の低下を抑制する。本開示のニューラルネットワーク処理装置１０は、複数の人工ニューロンが結合したニューラルネットワークＮ１に対して複数の入力データ４０を与えて、前記人工ニューロンから出力される複数の出力からなるベクトルを、複数の前記人工ニューロンそれぞれについて求める処理２２と、前記ベクトルに基づいて、同一又は類似の振舞いをする複数の人工ニューロンを選択し、選択された複数の人工ニューロンを統合する統合処理２３と、を実行するように構成されている。

学術雑誌等の編集委員・査読・審査員等

査読1件

学術雑誌等の編集委員・査読・審査員等

編集査読委員等,任期:1年

学術雑誌等の編集委員・査読・審査員等

編集査読委員等

公開講座・講演会の企画・講師等

まず，大量のデータ中から検索キーに類似したデータを高速に探索する問題にお いて，データ間の類似性をいかに定義するのかが極めて重要な問題であるという ことを実例を挙げながら示す。特に，検索キーベクトルに劣化（欠落）が含まれ る場合には，データの正規化方法と使用する類似性尺度によって，検索結果が大 きく変わることを示し，その理由について言及する。次に事例に基づく異常検出 において，主に用いられているSimilarity Based Modellingが，類似度をカーネ ル関数と同じであるとみなせば，非線形回帰の手法であるGaussian Processと本 質的に同じになるということを示す。,日付:10月31日

小・中・高校生を対象とした学部体験入学・出張講座等

後期入試の過去問「情報」を高校生向けに解説した．,日付:7月25日

小・中・高校生を対象とした学部体験入学・出張講座等

コンピュータの生い立ち，コンピュータを用いた計算で何が出来 るのか等について，現在高校生の皆さんが学んでいることを基 に説明を行った． ,日付:7月16日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「壊れたPCをRe活用した低価格Wi-Fiシステム技術を開発」
「世界初!Wi-Fi宝さがしin南紀田辺」開催
2010年11月6日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「宝探しイベント:廃棄PCで　和大研究室がシステム開発--きょう、田辺で　/和歌山」
2010年11月6日、和歌山大学システム工学部の学生たちが開発した廃棄パソコン(PC)による無線LANシステムを利用し、参加者が携帯端末で受信したクイズを解いて街を巡る宝探しイベント「Wi-Fi★宝さがしin南紀田辺」開催のご案内。
毎日新聞　2010年11月6日　地方版

研究成果に係る新聞掲載、テレビ・ラジオ出演

「携帯端末片手に街巡り　田辺の商店街で「宝探し」」
JR紀伊田辺駅周辺の商店街で6日、iPad(アイパッド)、iPhone(アイフォーン)など電子携帯端末を使って宝探しをするイベント「Wi-Fi(ワイファイ)★宝さがし　in南紀田辺」が開かれた。当日は端末の貸し出しもあり、家族連れや大学生など、県内外から約100人が参加した。
2010年11月11日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「和歌山大学、PCをRe活用したWiFiシステム技術を開発」
和歌山大学
壊れたPCを用いた無線広報システムとその応用事例に関するご案内
2010年10月20日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「和歌山大学、壊れたPCを活用した低価格のWi-Fiシステム技術を開発」
2010年10月20日

研究成果に係る新聞掲載、テレビ・ラジオ出演

最新機器で宝探し(11月6日)　和大生が田辺でまちおこし
和歌山大学システム工学部の和田俊和教授と同大学の学生や和大大学院の院生らでつくるプロジェクトチームが、まちおこしにつなげようと田辺市商店街振興組合連合会、田辺広域の第三セクターまちづくり会社「南紀みらい」とともに企画、運営する。

2010年10月29日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「和歌山大学、壊れたPCでも再活用できる低価格WiFi情報配信システムを開発」
　和歌山大学 システム工学部 和田研究室は10月20日、ハードディスクの故障などで廃棄されたPCのWi-Fi(無線LAN)通信機能を有効活用して、スマートフォンや携帯ゲーム機に独自情報を発信できるシステムを開発したことを発表した。
2010年10月20日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「【今週の Web ミミズク】和歌山大学で、壊れた PC が WiFi システムとして復活!」
和歌山大学のシステム工学部、和田研究室では、HDD の故障などで廃棄された PC の WiFi(無線 LAN)通信機能を有効活用して、iPhone などのスマートフォンやニンテンドー DS などの通信対応携帯ゲーム機などに情報を発信できる、低価格 WiFi システム技術を開発中だそうだ。
2010年10月23日

研究成果に係る新聞掲載、テレビ・ラジオ出演

「「田辺を遊園地に」　和大生らラリー企画で商店街訪問」
和歌山大学の学生や和大大学院の院生ら11人が7日、田辺市中心市街地の商店街を訪れた。多機能携帯電話などを活用して、参加者がクイズなどに回答しながら各商店を回る「ワイファイラリー」の企画を検討した。
2010年9月8日

公開講座・講演会の企画・講師等

ディジタル化された画像・映像は比較的容易に改竄することができる.ディジタルカメラで撮影された画像でさえ,写実性からの逸脱という意味で,改竄されているということもできる.本講演では,ディジタル画像,映像の改竄方法を紹介し,それらを見抜く方法を紹介する.,日付:2010.3

公開講座・講演会の企画・講師等

近年盛んに用いられるようになっている離散最適化の方法として,Graph CutsとBelief Propagationについて概説し,セグメンテーション,ノイズ除去など様々な用途に応用可能であることを示す.また,我々の開発した2分探索型Belief Propagationによる計算の高速化とFPGA上での高速化方法とその結果について述べた.,日付:2010.2

公開講座・講演会の企画・講師等

画像から対象を検出,追跡,識別するというタスクの中で,最近傍探索は数多く用いられる.最近傍探索アルゴリズムの概要と,検出・追跡・識別への応用について述べ,距離計算の違いによって画像検索精度が大きく変わるなどの最近の成果についても説明する.,日付:2010.1

公開講座・講演会の企画・講師等

最近傍探索の研究においては,「次元の呪縛」と呼ばれる現象が問題を困難にしてきた.この現象は,数学的に厳密に定義されていないが,「蓄積されるデータの分布の次元が一定の値を超えると,いかなるアルゴリズムでも全探索と等価になる」という現象である.最近傍探索研究の歴史は,これを解決するために行われてきたと言っても過言ではない.これまでに数々の研究がなされてきたが,次元の呪縛が解けたという報告はこれまでになく,近年はこの現象を回避するための「近似最近傍」を探索する研究が盛んにおこなわれるようになってきた.本チュートリアルでは,近似を含むものと含まないもの両方について高速な,日付:2009.11

公開講座・講演会の企画・講師等

モデルベースの対象表現法と事例ベースの対象表現法の方法と実際について述べ,画像からの歩行者検出タスクでは,近年盛んに研究がおこなわれるようになっている事例ベース対象表現法よりもモデルベース対象表現法が優れている場合があることを示した.これを通じて,流行ばかりを追求する研究の方法論を見直し,合目的的な方法論を採用すべきであることを訴えた.,日付:2009.6

国際交流事業

上記ワークショップの開催責任者,海外からの研究者招聘手続き,予稿集出版,他全て,相手国:米国,イギリス,スイス,中国,フランス,カナダ,オーストラリア

国や地方自治体、他大学・研究機関等での委員

国や地方自治体、他大学・研究機関等での委員,任期:2013/08/01～2014/07/31

国や地方自治体、他大学・研究機関等での委員

国や地方自治体、他大学・研究機関等での委員,任期:2013/08/20～2013/12/31

国や地方自治体、他大学・研究機関等での委員

国や地方自治体、他大学・研究機関等での委員,任期:2012/12/1～2013/1/31

学協会、政府、自治体等の公的委員

ICCV2011 の査読結果を集約し,論文採否の議論を行い,決定し,レポートを作成する業務。今年度は39本の論文について書かれた116通の査読結果を集約した。,任期:2011.4～2011.6

学協会、政府、自治体等の公的委員

日本の画像解析・認識研究の中心的組織であるPRMU研究会の運営を司る専門委員会の委員,任期:2010.5～2011.3

学協会、政府、自治体等の公的委員

日本国内でのコンピュータビジョン研究の中心的組織であるCVIM研究会の運営委員,任期:2010.4～2015.3

学協会、政府、自治体等の公的委員

ACCV2010のプログラムを決定するために,論文の採否を決定する委員会の委員である.,任期:2010.3～2010.11

学協会、政府、自治体等の公的委員

日本国内でのコンピュータビジョン研究の中心的組織であるCVIM研究会の運営委員,任期:2009.4～2010.3

学協会、政府、自治体等の公的委員

日本の画像解析・認識研究の中心的組織であるPRMU研究会の運営を司る専門委員会の委員,任期:2009.4～2010.3

学協会、政府、自治体等の公的委員

ICCVのプログラムを決定するために,論文の採否を決定する委員会の委員である.今回は日本開催であったため,全32人のうち日本から参加した委員は3名であった.,任期:2009.3～2009.9

学協会、政府、自治体等の公的委員

学協会、政府、自治体等の公的委員,任期:2008.3～2009.1

国や地方自治体、他大学・研究機関等での委員

Area Chair ,任期:2006.3～2007.10

国や地方自治体、他大学・研究機関等での委員

プログラム委員長 ,任期:2006.3～2007.8

学協会、政府、自治体等の公的委員

学協会、政府、自治体等の公的委員,任期:2003.3～2009.3

国や地方自治体、他大学・研究機関等での委員

運営委員, 幹事 ,任期:2003.3～2009.3

国や地方自治体、他大学・研究機関等での委員

論文誌編集委員 ,任期:2003.3～2009.3

国や地方自治体、他大学・研究機関等での委員

編集委員 ,任期:2003.3～2006.3