亀田 裕介

This paper proposes a spatially scalable image coding scheme based on a non-linear image resizing technique called seam carving. In this scheme, the original size image is reconstructed from the reduced size one with a help of side information including paths of the removed seams. To reduce the amount of the side information, evenly spaced multiple seams are removed simultaneously at the resizing stage.

In this paper, we propose a rotation invariant method for human detection from top view images. In the method, local features based on Histogram of Oriented Gradients (HOG) are defined based on polar coordinates. In addition, two dimensional histograms are calculated from pairs of neighboring pels to enhance robustness against small rotation.

We previously proposed a semi-automatic 2D-3D conversion method for still images. The method propagates depth information, which is partially given by the user as an input, to a whole image based on local similarity of image values. Therefore, the propagation process must be sufficiently fast for interactive operation. In this paper, we accelerate this process by using a kind of distance transform.

This paper describes a new intra prediction method which is designed to exploit inter-block correlations for DCT based image coding. By changing a coding process of DCT coefficients from block-wise order to coefficient-wise one, the method can benefit from interpolative prediction in almost all blocks.

We previously proposed a block-adaptive spatio-temporal prediction method for video coding. In this method, multiple predictors that minimize sum of squared errors are designed for each frame. However, such a least squares approach does not necessarily improve the coding performance. In this paper, we introduce a DCT domain soft shrinkage process to reduce the number of non-zero DCT coefficients to be encoded.

For the purpose of user interface or marketing, estimation of gaze directions by a noncontact measurement device has been studied in recent years. This paper describes an eigenspace-based method which estimates horizontal and vertical angles of the gaze directionfrom eye region images.

As an efficient way for archiving the existing video contents compressed by the old standard, we are developing a lossless re-encoding scheme for MPEG-1 video. In this paper, simplified probability models used for arithmetic coding of DCT coefficients are investigated to reduce computation time needed for our coding scheme.

This paper proposes an efficient lossless coding method for color quantized images each of which consists of an array of indices and corrsponding RGB values of representative colors. In the conventional methods, predictive coding techniques developed for lossless image coding are often applied to the array of indices. On the other hand, our method directly predicts a 3D position of each index in a RGB color space to accurately estimate its occurrence probability.

We previously proposed a new type of lossy image coding scheme whose compressed data can be losslessly converted into JPEG bitstreams. In this paper, perceptual image quality of our coding scheme is improved by taking account of spatial frequency characteristics of the human visual system. Results of our subjective quality assessment test demonstrate that perceptual quality of the reconstructed images is certainly improved.

We propose a stable computation scheme for variational scene-flow computation. Scene flow is a velocity field on a stereo image sequence and represents apparent motions and disparity changes. The stability condition of the proposed scheme is independent of images and regularization coefficients in the variational problem.

We previously proposed a lossy coding scheme based on recursive-type intra prediction and discrete sine transform (DST) for monochrome still images. In this paper, we extend the scheme to RGB 4:4:4 format color images. In order to exploit intra- and inter-color correlations, the already encoded color signals are simultaneously used in the prediction process.

Recently, the coding technology of multi-view video plus depth has been developed to implement 3D video applications such as Free Viewpoints Television. Because depth maps tend to have smooth value within an object and have strong edges between objects, it is efficient to encode depth maps by a region-oriented manner. In this paper, we introduce a segmentation method based on Voronoi diagrams to reduce the amount of information on the region shapes. Moreover, we propose a plane approximation to the depth values in each region.

In the previous work, we proposed a lossless coding method which dynamically models probability density function (PDF) of image values pel-by-pel. In this method, the PDF model is defined as sum of the generalized Gaussian functions (GGFs) whose peak positions are given by examples collected from causal neighbors. However, the shapes of the GGFs were fixed for each image. In this paper, we introduce a technique of context modeling to adaptively control the respective shapes.

Directional intra prediction currently used in H.264/AVC and H.265/HEVC is unsuitable for curved textures because predicted values are always extrapolated along straight lines in a given direction. In this paper, we propose a new prediction technique which enables interpolative prediction along circular curves defined by two kinds of parameters.