Stereo Correspondence Using ART Neural Networks: Algorithms and Parallel Implementation.
August 18, 1997
One of the long standing problems in passive stereo vision is the construction of an accurate range map from only two images that provide two views of the same 3-D world scene. It amounts to identifying corresponding pixels between the two images (left and right) that are associated with the same point in the 3-D world. In this work we are introducing the use of an Adaptive Resonance Theory (ART) neural network as an associative memory device for solving the stereo correspondence problem. Using a multi-pass approach, in which the vigilance parameter value is gradually reduced, it is possible to increase the density of matched points. On the other hand, by imposing a requirement of close agreement between disparity values in a neighborhood, false positive matches are eliminated. At the end, an accurate and reasonably dense range map is obtained, to the extend that it allows 2.5-D reconstruction of the schene via standard interpolation methods. The scheme has been tested on random dot stereograms, virtual world scenes generated by computer programs (where the ground truth is known) as well as real-world scenes. In all cases the the 2.5-D reconstructions are shown to be quite realistic. Furthermore two parallel ring implementations of the ART network have been introduced and a complete scalability analysis has been performed and verified on a MIMD Transputer based multiprocessor platform.
Prof. Dana Brooks
Prof. Elias S. Manolakos (thesis advisor)
Dr. Homer Pien (Draper Lab)