Recovering depth from stereo using ART neural networks

One of the long standing problems in passive stereo vision is that of constructing a range map using only two images providing two views of the same 3-D world scene. It amounts to identifying pairs of corresponding pixels, one pixel in each image, that are associated with the same point on the 3-D world. We are introducing ART-1 neural networks as a primitive capable of addressing the stereo correspondence problem. Using a multi-pass approach it is possible to increase gradually the density of matched points, while at the same time false matches are filtered by requiring close agreement between disparity estimates in a neighborhood. At the end, a reasonably dense disparity map is obtained, to the extent that it allows the reconstruction of the scene using interpolation methods. Our scheme has been tested on stereograms, virtual world scenes generated by computer programs (where the “ground truth” is known) as well as real-world scenes. In all cases the scene reconstructions are shown to be quite realistic