Kenichi Arakawa

Appearance based pose estimation of 3D object using support vector regression

Several methods for estimating the pose of a 3D object from its appearance have been proposed. The parametric eigenspace method is typical of such methods. One key disadvantage of this method is that storage requirements explode when the degree of freedom is increased. In this paper, we propose a method of suppressing this increase in storage requirements by describing the relationship between an image and a pose as functions. Pose estimation functions, which keep the generalization ability high even if the storage requirements are small, are obtained by using support vector regression. Experimental results show that the proposed method can compress the storage requirements to just 1/100 of that needed by the parametric eigenspace method.

Kanji recognition in scene images without detection of text fields - robust against variation of viewpoint, contrast, and background texture

With the goal of indexing scene images, we propose a novel recognition method for Kanji characters captured in scene images. Our method scans multi-resolution images and classifies clipped regions with recognition dictionaries generated by learning a large amount of partial patterns of characters with large geometric transformation. The problem of scanning time, which tends to be unpractically long, is solved by using multi-compression coarse-to-fine scanning, and by detecting peak points after coarse searching. Despite the wrong results generated in the background, our method well supports image retrieval since it uses the regular spacing of characters. Experimental results show that this recognition method recognized characters at the rate of 82%. Precision was 84% and recall was 64% for image retrieval.

Tangible search for stacked objects

The goal of Tangible Search is to more effectively support the user in physically locating one of a number of stacked objects. It consists of two operations - automatic logging of stacked objects and direct annotation; image processing is used to determine the heights of the stack and of the users finger. Tangible Search offers stable and accurate 3D analysis since it uses our previously proposed method. It employs a single camera with a compound half-mirror; this configuration also allows the top and side views of the stack to be captured simultaneously. Our approach is to make it easier for the user to handle stacks of items; it will enhance the tabletop metaphor for intuitive interaction in real-world environments where stacks are very common.

Motion and shape recovery based on iterative stabilization for modest deviation from planar motion

We describe an iterative stabilization method that can simultaneously recover camera motion and 3D shape from an image sequence captured under modest deviation from planar motion. This technique iteratively applies a factorization method based on planar motion and can approximate the observed image points to the 2D points projected under planar motion by stabilizing the camera motion. We apply the proposed method to aerial images acquired by a helicopter-borne camera and show better reconstruction of both motion and shape than Christy-Horauds perspective factorization. Moreover, we confirm that the reprojection errors calculated from the recovered camera motion and 3D shape are very similar to the optimum results yielded by bundle adjustment

Estimating Anomality of the Video Sequences for Surveillance Using 1-Class SVM

We have proposed a method to detect and quantitatively extract anomalies from surveillance videos. Using our method, anomalies are detected as patterns based on spatio-temporal features that are outliers in new feature space. Conventional anomaly detection methods use features such as tracks or local spatio-temporal features, both of which provide insufficient timing information. Using our method, the principal components of spatio-temporal features of change are extracted from the frames of video sequences of several seconds duration. This enables anomalies based on movement irregularity, both position and speed, to be determined and thus permits the automatic detection of anomal events in sequences of constant length without regard to their start and end. We used a 1-class SVM, which is a non-supervised outlier detection method. The output from the SVM indicates the distance between the outlier and the concentrated base pattern. We demonstrated that the anomalies extracted using our method subjectively matched perceived irregularities in the pattern of movements. Our method is useful in surveillance services because the captured images can be shown in the order of anomality, which significantly reduces the time needed.

Active Viewpoint Control for Shape from Occluding Contours

The contour of an object contains a lot of useful information. For reconstructing the shape of the objects surface, a sequence of contours extracted from object images taken from controlled viewpoints is often used. However, there are many factors that cause errors in reconstruction. One factor — the intervals between the viewpoint positions — can produce serious errors. Taking images at finer intervals decreases these errors, but increases the time, as well as other costs, of taking and processing images. This paper introduces a method for controlling the intervals of viewpoint positions, which enables the errors caused by those intervals to remain within a previously decided fixed limit and can reduce the required number of viewpoint positions.

Online anomal movement detection based on unsupervised incremental learning

We propose an online anomal movement detection method using incremental unsupervised learning. As the feature for discrimination, we extract the principal component of the spatio-temporal feature by incremental PCA. We then detect anomal movements by an incremental 1-class SVM. In order to use principal component as the feature for discrimination while supporting incrementation of the subspace, we modify the SVM kernel function to take account of the difference in distance scale between the principal component feature vectors and that of the feature vectors after the subspace is incremented. This allows us to efficiently conduct the relearning process even though the dimension of the original input spatio-temporal feature is high. Experiments show that anomal scenes can be detected without the cost of preparing a lot of labeled data for preliminary learning.

Shape from contour using adaptive image selection

The “shape-from-contour method” reconstructs the 3D shape of the surface of an object by extracting its contour from each of a series of successive images of the object. This can be realized by using a CCD camera, and is a relatively accurate method of obtaining environmental information. However, to obtain an accurate result, many images must be processed. Therefore, it is important to select the images depending on their effect on the final result, particularly for high-speed processing. This paper proposes an adaptive image selection (AISE) method which depends on the required accuracy. The method has been applied to objects having various cross-sectional shapes. The errors in shape reconstruction and the number of images required in the conventional method and the proposed method are compared. The experimental results show that the proposed method requires fewer images than the conventional method, particularly when the surface curvature of the object has large variation. The theoretical relationship between the accuracy of the reconstructed shape and the number of images required is also derived.