[in Japanese]

View full abstract

[in Japanese]

View full abstract

A method for measuring 3-D objects with surfaces composed of various reflective regions is developed in which the 3-D coordinates of the object sections are obtained by triangulation with the light-section method by using a slit-light source and a TV camera. A slit-light source in which the lighting intensity is controllable and a TV camera with an expanded dynamic range are employed to reduce the influence of uneven reflective surfaces. The TV camera expands its dynamic range by combining multiple images captured at different shutter speeds. A trial 3-D shape measurement system is produced using this method, and its effectiveness is confirmed by experiments.

View full abstract

In order to synthesize intermediate-view images of arbitrary viewpoints between a pair of left and right view images, we propose two methods for disparity estimation with discontinuity at object contours. Both methods have four steps : (1) Two initial disparity maps based on left and right images are estimated with block matching. (2) Ill-corresponding areas in both of the initial disparity maps are detected using 'confidence evaluation'. The object contours are, then, detected by (i) Snakes method or (ii) Automatic Contour Detection (ACD) method. (3) The disparity maps are reestimated with discontinuity at the object contours. (4) The reestimated disparity maps are used to synthesize two intermediate-view images based on both view images and these images are integrated into the final intermediate-view image. Our methods can estimate the disparity of occluded areas with discontinuity at the object contours. Experimental results show our methods yield better quality intermediate-view images near the object contours than do ones using the initial disparities.

View full abstract

This paper describes a new method for tracking objects in 3D motion from image sequences by using constraints in the scene. Many methods based on the gradient scheme have been proposed for tracking objects. However, these methods have a serious drawback in that tracking error sometimes increases during a long image sequence due to various noise. The proposed method enables us to track the objects over a long time by the gradient-based constraint with some scene constraints; that is, constraints on position, velocity, and acceleration in the scene. Experiments with real image sequences support the precise tracking over a long time.

View full abstract

It is difficult for us to detect a disparities from stereoscopic images This paper presents the solves of stereo matching problems that is 2 camera reflecting conditions setting problem Usually, a stereo matching methods need a fine setting stereoscopic camera But, it is so difficult to set the camera conditions Therefor, we developed new fitting method for stereo images The principle is very simple If stereo images are fitted, the disparities detecting ratio that is made by probability becomes high So, we selected the highest detecting ratio conditions i e the amplifier gain conditions or offset value conditions, and the Y-axis aberration or camera rotation, etc The detected disparities are about 120 pixels using a stereoscopic camera that has 640 (H) ×480 (V) pixels, a 20° angle of view and a 900 mm subject distance

View full abstract

This paper proposes a method of estimating multiview camera parameters for cases in which the structure of objects is unknown. The proposed method uses multiple rectangles in the 2-D images captured with the same focal length as the keys to estimate camera parameters. In this method, the 3-D coordinates of rectangles are first estimated from all images by using the characteristic of rectangles and the least squares method. Then, all camera parameters are estimated by using the estimated 3-D coordinates of rectangles and the least squares method. Finally, all parameters, that is, 3-D coordinates of rectangles and camera parameters, are corrected to minimize the errors in all images. Experimental results show that the proposed method gives good estimations of the multiview camera parameters required to obtain ray information for cases in which the shape of objects is unknown. This means that ray information can be acquired by using one ordinary camera instead of a special camera controlled by computer.

View full abstract

In this paper, we propose a novel data compression scheme based on ray-space coding for 3-D spatial information. By using a spatially uniform sampling method for the sampling of the ray-space and estimating the amount of ray-space data, we can efficiently reduce the redundancy of the ray-space data through disparity compensating predictive coding. From this viewpoint, we propose a novel predictive coding scheme which utilizes the geometrical relationship between orthographically projected view images. In our experiment, a compression ratio of 1/100 with PSNR 27 dB is obtained for computer-generated ray-space data.

View full abstract

This paper describes a new data compression method for spatial partitioning representation which gives a higher data compression ratio than conventional methods. To obtain high data compression, we introduce the Network model, which performs fast geometric operations for spatial partitioning representation. The Network model represents a shape by using only voxels on the surfaces. These voxels connect with each other to make a network structure. The model compresses data simply for spatial partitioning representation and has a better compression ratio than the Octree model. Moreover, the proposed method can eliminate a large quantity of voxel from Network model. The resulting compression ratio is better than the three-dimensional runlength approach. Experimental results show that the method is effective for compressing data for spatial partitioning representation and achieves a high compression ratio.

View full abstract

We propose a method to generate an arbitrary view based on five reference views. We attempt to synthesize a new view with the best attainable quality given limited information. The method is based on estimation of a depth map and successive image mapping : transformation of the depth map, guessing the depth map for the uncovered area through a simple directional interpolation, and back-mapping using the filled depth map with a reasonable check. The uncovered area due to a view change can be handled on the basis of the observable viewpoint. We confirmed that an acceptable image quality could be efficiently synthesized for an arbitrary viewpoint within a scope.

View full abstract

This paper examines shooting methods for natural 3-D images in 3-D program production. Although toed-in and parallel camera configurations are both possible, toed-in configuration is usually used because shooting and viewing conditions can be easily set to obtain the desired 3-D design. However, this camera arrangement has been shown to bring about inconsistencies between the depth information from perspective of the lenses and the depth information from binocular parallax, leading to size distortion called the Puppet Theater Effect. In contrast, geometrical calculations show that the parallel camera arrangement does not cause such inconsistencies under specific shooting and viewing conditions called “orthostereoscopic conditions.” The possibility of representing 3-D images using this method is studied through 3-D program production. Subjective evaluation tests show that the 3-D images shot under orthostereoscopic conditions look more natural than those shot by toed-in camera configurations at any display size.

View full abstract

This paper describes an image processing and generating method which replaces a human face in motion pictures with another person's face. First, we set the positions of four feature points of the facial region of person A on the first frame of a motion picture M₁ manually. Their positions on the following frames are determined almost automatically to obtain the three dimensional motion of person A's face. Second, the facial image of person B is extracted from a motion picture M₂, and mapped to the surface of the 3-D patches pre-generated from 3-D data of person B's face. Third, for each frame in M₁, the 3-D patches are transformed according to the direction of person A's face and synthesized with it to generate the frame of a synthesized motion picture M₃. Finally, a post-processing is executed to make M₃ natural. Algorithms of the proposed method are described with several experimental results.

View full abstract

We have developed an image conversion system that converts two dimensional (2D) images into stereoscopic images (3D images) with the modified time difference (MTD) method. This method allows us to convert ordinary 2 D images into binocular parallactic 3D images according to the detected movements of objects in the images. Furthermore, we have developed an algorithm that controls the binocular parallax to enable the creation of smooth 3D images. With this system and algorithm, automatic realtime image conversion can be realized. We implemented the MTD method in a single LSI, which made the 2D/3D conversion board very compact.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract