[in Japanese]

View full abstract

Various shaped screens have bee recently used for immersive projection displays. The ability to project a non-distorted image on a curved screen from arbitrary viewpoints is highly desirable. However, in previous techniques, the shape of the screen has been restricted, and we required geometrical information about screens and projectors. In this paper, we descrive a non-distorted image projection method for an immersive projection display with a curved screen. This method contains two approaches : the first is a precise geometry and color correction for preprocessing with a digital camera, and the second is a real-time geometry correction with real-time 3D rendering techniques. These methods need only geometrical information about the screen, and achieve a high quality image correction in real-time. We also used this method in the actual immersive projection display, and evaluated its effectiveness in some experiments.

View full abstract

In this paper, we present a new method for extracting the 3D graph structure of gallery regions (worm-eaten areas) from x-ray CT images of wooden statues of Buddha. The extraction and analysis of these gallery regions by image processing is indispensable in the restoration of these historic and precious statues. Our new method consists of two parts : extraction of the initial 3D graph structure and its correction. Because the local shape of the statues and of the gallery regions can be similar, extraction of the gallery regions can create errors in the initial 3D graph structures. The second part of our method satisfactorily corrects both extraction errors and unsuitable extractions based on their local and global features. Our method can extract the 3D graph structures of gallery regions with high accuracy and can also be used to visualize these regions using extracted graph data.

View full abstract

This paper proposes an animation method for crowd behavior in three-dimensional space. In this method, each movement model in the computer generated virtual reality space has a movement vector and personal space to avoid collisions with other movement models. When a movement model enters the personal space of an another movement model, the movement model avoids a collision with the incoming models by a calculation of the movement vector. The uniqueness of the proposed method is that real time behavior animation in crowds can be realized by a simple geometric calculation, and the general idea of personal space can be applied to fast moving objects. Our experimental results demonstrate that this method is effective in real time animation.

View full abstract

This paper presents a novel method for arbitrary view synthesis that allows viewers to virtually fly through in real soccer scenes. Multiple cameras situated around a stadium capture the action, and images shot from arbitrary viewpoints are generated by viewpoint interpolation, using projective geometry between neighboring cameras. The scenes are segmented according to their geometric properties.Dense correspondence matching between real views automatically occurs by applying projective geometry to each region. Superimposing the intermediate view images, synthesized in every region completes the virtual views of the entire soccer scene. Camera calibrations can be reduced, and correspondence matching requires no manual operation, allowing the proposed method to be applied to dynamic events in a large space. In addition to the view synthesis technique, we introduce an application for fly-through soccer match observation. This technology will lead to the creation of new media that can be applied to a variety of entertainment, from concerts to sporting events.

View full abstract

In this research, we address the problem of finding the best correspondence in matching algorithms in the edge area and we propose offset-block matching for ray-space interpolation. Conventional block matching methods are pixel-based, block-based and hierarchical methods. The pixel-based matching method performs well and satisfactorily detects correspondence in the edge area, when the search area is small. Block-based matching methods detect better than pixel-based methods where is a large disparity, however the detecting of correspondence in the edge area is not as accurate. The hierarchical block matching method performs relatively well in all areas. The proposed offset-block matching however combines pixel-based and block-based matching methods by limiting the search area and obtains better results than conventional matching methods. The proposed method has simultaneous advantage of both the pixel-based and the block-based matching methods. Furthermore, this method is suitable for distributed architecture of a video rendering system like FTV, when we need real-time interpolation.

View full abstract

To display 3D images on a computer-generated hologram, you need to have a very high resolution hologram-recording system. We developed a CD-R (Compact Disk-Recordable) drive to record CGH data on CD-R disks, that have a high resolution and a dot size of 1.5μm. By using this drive, it is possible to make holograms inexpensively in a short-time without the need for special environments, such as dark rooms, clean rooms, etc. The drive realizes desktop holograming that can handle all CGH process including model making, hologram calculations, and hologram recordings on a personal computer. In this paper, we describe the system and discuss the hologram-calculation method adapted for CD-R disks.

View full abstract

A Walsh transformation pair technique is used to analyze the shape features of linear drawings, providing superior shape extraction for angular features. This method employs real number calculations only, but at lower computational cost than previous techniques. Difficult shape deformation (transformation of squarish shapes) can be performed easily using this method. Feature analysis and shape deformation for linear drawings of arbitrary shapes are comparable to previous methods, and the transformation of various shapes can be achieved by combining the proposed method with previous methods.

View full abstract

In this paper, we propose a new 3D modeling method called DSFS. This method combines two methods, shape from silhouette and shape from shading, to compensate for their faults while retaining their advantages. The DSFS method models the object accurately and quickly, and, at the same time, it solves the concave surface modeling problem what was difficult for conventional shape from silhouette methods. Compared to other shape from shading methods, the iterative stage of DSFS is more stable and more accurate, without sensitivity to the experimental environment. The effectiveness of DSFS has been confirmed by the experimental results though computer simulation.

View full abstract

If a circular mask is put on a human face to cover its main parts (an eye, a nose and a mouth). the content is invariant with the rotation in image planes. This invariance can be computed by the magnitude spectrum of Fourier transformation with polar coordinate systems. The correlation and rotation angle compared with a trained face model can further be computed based upon phase correlation method. However, the computation is necessary to carry out polar coordinate transformation at every position of an image for every face-finding processing when we use a normal rectangular input, lattice. In this paper, we take advantage of the characteristics of hexagonal lattice, sample an input image of rectangular lattice by hexagonal lattice, approximate the circular mask with a hexagonal one, and find a face independent of its rotation in an image plane.

View full abstract

Most recent video coding methods like MPEG, are types of hybrid coding that use motion compensation (MC) and discrete cosine transform (DCT). In DCT, because images are partitioned into macro blocks, the efficiency decreases for a very low bitrate. We propose a new video coding method that uses matching pursuits (MP). MP represents a signal approximately by using a dictionary to obtain a high coding efficiency for a very low bitrate. In this paper, we propose a matching pursuits coding method that uses a dictionary suitable for the sub-band frequency characteristic for motion compensated prediction error images, and will demonstrate how this method can improve coding efficiency.

View full abstract