Stereo correspondence based on curvelet decomposition, support weights, and disparity calibration

Abstract

A novel multiresolution analysis-based stereo matching method using curvelets, support weights, and disparity calibration is proposed. By introducing curvelet decomposition, we obtain the curvelet coefficients in different scales and orientations, and the image points can be better described and represented by these coefficients. By using support weights, the fattening effect suffered by previous methods is reduced. As a result, false matches are reduced greatly and overall accuracy is increased. Disparity calibration smoothes the disparity map and removes the remaining outliers to further improve accuracy. The proposed method is verified and compared extensively with state of the art methods, and good results and improvements are achieved.