Rejecting Wrong Matches in Stereovision

Block matching methods in stereo vision must de- cide when two pixels match, and usually do it by thresh- olding a similarity measure between two blocks. This paper proposes a more elaborate statistical block matching method ensuring that the detected block matches are not likely to have occurred "just by chance". The method is based on the generation of a simple but faithful statistical background model for image blocks. The ensuing rejection/acceptation process uses an a contrario method that computes a false alarm numbers, and guarantees that on average not more than one wrong block match occurs. To evaluate the perfor- mance of this method, its application range will be first es- tablished. Two subsidiary matching criteria prove necessary: First, the fattening effect inherent to block matching meth- ods is avoided by comparing only blocks that do not meet the image line segments on which the depth is discontinu- ous. Second, the stroboscopic errors due to repetitive shapes are avoided by a SIFT-like threshold. The method is tested on three kinds of data: urban aerial scenes with low baseline, classic benchmark examples, and a close range stereo pair. All confirm that a careful three steps error analysis yields a dense disparity maps with 0.3% wrong matches and less. The presented technique is tested with block matching, but can be applied as an after check to any other stereo matching method.