We are developing a teleconferencing system that offers realistic sensations, as an application of Virtual Reality, and are researching ways to synthesize real objects and stereoscopic images. At present, there is not much information on the fusional limits in the situation where the real objects and stereoscopic Images are synthesized. We will provide some data on fusional limits for such a situation, and discuss a design method of displaying stereoscopic images.
A neural network model for extraction of binocular disparity Is proposed based on knowledge of visual system In physiology. The abilities of extraction of binocular disparity are compared among four proposed cell models corresponding to binocular simple cells in vsual area 1. As a result of computer simulations using a random-dot, stereogram, it is shown that a cell model which realizes multipication between outputs of a left and a right, cell on the lower layer has best performance of extraction of binocular disparities. It is also shown that cell models with inhibitory connections have better performance than a cell model without inhibitory connections.
An objective picture quality scale for high definition still pictures coded by the JPEG Scheme is studied. A scale is derived from two types of weighted root-mean-square-errors, those in flat areas and non-flat areas. The two areas are determined by the standard deviation in the sub-blocks of the original picture. The scale value show a good linear correlation with subjective picture quality.
Our aim is to evaluate the relationship between the mechanism of processing visual information and eye movement. To examine the characteristics of the acquisition of visual information at a fixation point, eye movement was analyzed while the subjects counted circles embedded in distractors. In our new approach, distractors that were included within a certain area around the fixation point were changed to circles, immediately after the subjects fixated on them. As a result, fixation duration of 100 - 150 ms was required to correctly count circles, before the distractors of the fixated area of 6.7 degrees were changes to attractors.
The various edge detectors contain the difficult and conflictive problems that the fine resolution filter detects a lot of noise and that the coarse resolution filter causes ambiguities of localization. The multi-scaled edge integration is one of the most effective method to solve these problems. However conventional multi-scaled methods aims to solve only one of these problems. We propose a new multi-scaled edge integration method which uses weighted summation of Gaussian filtered edge patterns. This method realizes a noise-reduced and good localized edge integration. In this paper, the advantages of our method are theoretically explained and demonstrated with some experiments.
There are some studies investigating salient features in face recognition, but we do not know the saliency of facial features in somewhat ambiguous retrieval tasks such as retrieving "similar" faces or "favorite" faces. Different cognitive operations such as image creation may be involved in the favorite - face retrieval, and the required similarity judgment may be different between the similar - face retrieval and the standard face recognition. In this study subjects were asked to retrieve either similar faces or their favorite faces from the database of approximately 60,000 different line - drawn faces. The results showed that in the favorite - face retrieval task most of the individual features and feature relations were of similar importance but that in the similar - face retrieval task three individual features were of marked imortance.