Stereo matching using a modified efficient belief propagation in a level set framework

Stephen Goyer Rogers


Stereo matching determines correspondence between pixels in two or more images of the same scene taken from different angles; this can be handled either locally or globally. The two most common global approaches are belief propagation (BP) and graph cuts. Efficient belief propagation (EBP), which is the most widely used BP approach, uses a multi-scale message passing strategy, an O(k) smoothness cost algorithm, and a bipartite message passing strategy to speed up the convergence of the standard BP approach. As in standard belief propagation, every pixel sends messages to and receives messages from its four neighboring pixels in EBP. Each outgoing message is the sum of the data cost, incoming messages from all the neighbors except the intended receiver, and the smoothness cost. Upon convergence, the location of the minimum of the final belief vector is defined as the current pixel's disparity. The present effort makes three main contributions: (a) it incorporates level set concepts, (b) it develops a modified data cost to encourage matching of intervals, (c) it adjusts the location of the minimum of outgoing messages for select pixels that is consistent with the level set method. When comparing the results of the current work with that of standard EBP [11], the disparity results are very similar, as they should be.