1996 Volume 116 Issue 7 Pages 826-834
A neural network model with population coding for extracting binocular disparity is proposed. The population coding is expected to represent binocular disparity with hyperacuity. In order to solve the false target problem in stereo vision, some constraints, such as compatibility, uniqueness and smoothness, are effective. In general, the uniqueness constraint is realized by selecting an appropriate binocular disparity among some candidates. Accordingly, it is difficult that the uniqueness constraint is realized in a system with population coding.
The presented neural network model realizes the uniqueness constraint and population coding simultaneously. The model also interpolates binocular disparities at pixels where features to extract binocular disparity do not exist. The model is based on physiological findings.
It is shown by computer simulations that the model is able to extract almost correct binocular disparities in terms of center of gravity of distributed binocular disparities from random-dot stereogram forming a sphere and a plane, as well as two planes. In other words, the model is effective to extract binocular disparity from curved objects by means of population coding.