Learning and Feature Selection in Stereo Matching

摘要

We present a novel stereo matching algorithm which integrates learning, feature selection, and surface reconstruction. First, an instance based learning (IBL) algorithm is used to generate an approximation to the optimal feature set for matching. Second, we develop an adaptive method for refining the feature set. This adaptive method analyzes the feature error to locate the sources of mismatches. Then the search through feature space for maximizing the class separation function is guided by eliminating the sources of mismatches. Third, we introduce a method for determining when apriori knowledge is necessary for discriminating between the correct match and the sources of mismatches. If the apriori knowledge is necessary then we use a surface reconstruction model to discriminate between match possibilities. We performed comprehensive comparison of our algorithm and a traditional pyramid algorithm on a wide set of real images. Finally, extensive empirical results of our algorithm based on real images are presented.