|M.Sc Student||Hendel Tal|
|Subject||A Solution to the Stereo Correspondence Problem Using|
Constraints Imposed by Physical Surfaces
and the Gradient Disparity Limit
|Department||Department of Biomedical Engineering||Supervisor||Professor Moshe Gur|
|Full Thesis text|
It has long been realized that the first stage in the attainment of stereoscopic depth is the establishment of correspondence between matching features in the two retinal images. The brain seems to solve this correspondence problem (or matching problem) easily, best demonstrated by the vivid sense of stereoscopic depth that we get from random-dot stereograms, but efforts to understand how the brain achieves this feat have shown the problem to be surprisingly difficult. This work offers a solution to the stereo correspondence problem that is accurate, robust and biologically feasible.
Firstly, it is suggested that prior to the matching process itself a preceding stage ought to take place. This preceding step, which is dubbed de-convergence, serves to cancel the effects of vergence and version on the two retinal projections. It is shown that if the retinal images go through a de-convergence process, vertical disparities are eliminated from them thus allowing the search for corresponding points to be conducted only along horizontal lines. A very desirable outcome of the de-convergence process is that the laws of visual direction emerge as a natural corollary of the need to fuse corresponding points.
Secondly, a detailed analysis of the constraints imposed on the "behavior" of points in the retinal projections is performed assuming that points tend to lie on continuous smooth surfaces, rather than be randomly positioned in space. The outcome of this analysis is the idea that in order to find its corresponding point, every point has to define its own Signature, which is a description of the point's neighborhood in several levels of preciseness, and then look for the point in the other retinal projection with the most similar Signature. This Signature idea is then developed into an algorithm. An implementation of the algorithm is tested on various textured surfaces with excellent results. It is further shown that algorithm integrates the two rivaling notions of binocular fusion- Panum's fusional areas and the disparity gradient limit. Moreover, the recognition of monocular points is incorporated naturally into the algorithm without the requirement to add a special stage for this purpose, a requirement that appears in some previously suggested correspondence algorithms. Lastly, it is argued that the new algorithm is consistent with psychophysical and biological observations.