We have developed a SIMD-type neural-network processor (NEURO4) and its software environment. With the SIMD architecture, the chip executes 24 operations in a clock cycle and achieves 1.2 GFLOPS peak performance. An a...
详细信息
We have developed a SIMD-type neural-network processor (NEURO4) and its software environment. With the SIMD architecture, the chip executes 24 operations in a clock cycle and achieves 1.2 GFLOPS peak performance. An accelerator board, which contains four NEURO4 chips, achieves 3.2 GFLOPS. In this paper we describe features of the neural network chip, accelerator board, software environment and performance evaluation for several neural network models (LVQ, BP and Hopfield). The 3.2 GFLOPS neural network accelerator board demonstrates 1.7 GCPS and 261 MCUPS for Hopfield networks.
The authors aim at capturing 3D human motion from stereo images in real time without any attached markers. Their strategy comprises two tasks: identification of body parts and motion assignment to them. They used a pa...
详细信息
The authors aim at capturing 3D human motion from stereo images in real time without any attached markers. Their strategy comprises two tasks: identification of body parts and motion assignment to them. They used a pair of artificial retina chips for implementing stereo vision. They introduced three algorithms. The first algorithm is 3D motion extraction based on optical flow analysis, and they showed its performance with 3D CG. The second algorithm is to localize an image region including a human head by frame subtraction. The final algorithm is to extract the head rotation in depth. They successfully discriminated four kinds of head rotations. To demonstrate the performance, they built a demo system and realized that each rotation could be reflected on a head CG model.
We introduce eigenspace analysis of image projections and demonstrate that it yields approximately linear representations of facial rotation in the directions up-down and left-right, respectively. The approach uses un...
详细信息
We introduce eigenspace analysis of image projections and demonstrate that it yields approximately linear representations of facial rotation in the directions up-down and left-right, respectively. The approach uses unsupervised learning-the representation is established even without explicit knowledge of the actual face pose. The method is computationally very inexpensive, as it uses only image projections, a very low-dimensional image representation, and a small number of principal components. In addition, the approach allows us to make effective use of the built-in image projection functions of our artificial retina chips. For a number of applications the method offers thus a fast alternative to more precise and more general, but also more complex methods for determining facial pose.
We introduce wavelet-domain principal component analysis and show that it overcomes some of the limitations of space-domain principal component analysis without introducing computationally expensive processing steps. ...
详细信息
We introduce wavelet-domain principal component analysis and show that it overcomes some of the limitations of space-domain principal component analysis without introducing computationally expensive processing steps. We argue that distance measurement in the wavelet domain is psychovisually more appropriate for judging facial similarity than distance measurement in the space-domain, and build binary facial trees using principal component analysis in the wavelet domain. Compared to caricaturing of space domain similarity trees, caricaturing these trees in the wavelet-domain results in better feature alignment and thus sharper and more credible images. Nonlinear preprocessing makes the approach robust with regard to both global illumination changes and local illumination fluctuations that vary slowly in the spatial domain.
暂无评论