Belief Propagation With Message Passing Constraint In Stereo Matching

Stereo matching could create a dense disparity map from related views and be of great importance for applications of 3DTV, machine vision navigation and virtual reality etc. Stereo matching based on belief propagation(BP) globally optimizes the energy function and produces an accurate depth map with low computational complexity by propagating the belief of disparity labels. However, the low confident labels of occluded nodes and the depth discontinuity at object boundaries may substantially reduce the accuracy of the disparity map.This thesis focuses on the constrained message passing in the cases of regions occluded and the depth discontinuity in belief propagation.1. An occlusion detection method with multiple cost cross-validation and a subsequent consistency-cost replacement is proposed to eliminate the harmful effects caused by occluded nodes. Firstly, the consistency costs is obtained by summing up absolute differences(SAD) in a local window and formed a score function with Gaussian weights assigned. Secondly, determine the occluded nodes by a threshold. The replaced cost of an occluded node would be propagated in the successive iterations. Experimental results show that the bad matching rate of BP with occlusion detection(O-BP) is about 20% less than the regular BP while the time-consuming remains almost the same.2. A message passing suppression method to stop the incorrect BP at object boundaries(Segmentation based Belief Propagation, S-BP) is suggested. Furthermore, a Multiscale S-BP(M-S-BP) aimed at improper segmentation is introduced. The color based Mean Shift(MS)segmentation boundaries stop the message passing across two segmented regions, while M-S-BP suppresses the propagation in different levels in hierarchical BP to ensure regular propagation in regions with smooth disparity in case that unreliable single scale segmentation occurs. Experimental results demonstrate that the percentage of bad matching pixels of S-BP is about 25% less than the regular BP, and M-S-BP lower the bad matching rate further.