M. Ezzati, M. D. Levine Binocular stereo vision is a biologically motivated approach that uses two slightly different views of a scene to extract information about its three-dimensional properties. The two underlying principles of our approach to stereopsis are parallel computation of binocular disparities and the use of the resulting disparity map for image segmentation. The method divides the two images of the stereo pair into small sections and obtains initial estimations for the disparities of all such sections. The devision of the image pair is motivated by data representation in the ocular dominance columns of the primary visual cortex where information from the left and right eyes are represented in the form of interlacing image ``patches''. The algorithm uses cepstrum, a method traditionally used in echo detection and provides a local estimation of binocular disparity between corresponding patches. We study cepstrum and its properties and use the developed insight in two ways. First the properties of cepstrum are used to offer improvements to the initial disparity estimation stage. Second the knowledge of the performance of cepstrum is used to provide an interpretation of the initial disparity map and motivate a method for its refinement. In the next stage, we refine the initial disparity estimations using neighbouring disparity information. We employ a modified median filtering scheme for the refinement stage. Finally, based on the characteristics of the methods used in the disparity estimation and refinement stages, we illustrate that the overall disparity map is suitable for figure-ground separation. Our approach to stereopsis is different from traditional stereo algorithms, which use simultaneous search and optimization, in the following ways:

- The initial disparity estimation is a local computation which can be implemented in parallel. This provides two fundamental advantages for real-time systems. The first advantage is increased computational efficiency by allowing parallel estimation of stereo disparity instead of the more common iterative and sequential search. The second advantage of the algorithm is a fixed running time which is independent of image properties. This is due to the elimination of optimization from the search process.
- Naturally, the depth map contains information about the three-dimensional properties of the surfaces in the scene. However, its most salient feature is the discontinuities in distance which mark the boundaries between various surfaces and objects in the scene. The overall disparity map is used for differentiating between surfaces that are located at different distances and not for precise surface reconstruction. Using stereopsis for figure-ground segmentation, rather than surface reconstruction, eliminates the need for camera calibration which is essential in exact depth calculation. Therefore the approach is well-suited to active vision systems in which the cameras are in constant motion.

Mon Apr 7 12:54:24 EDT 1997