Disparity Filtering with 3D Convolutional Neural Networks

Stereo matching is an ill-posed problem and hence the disparity maps generated are often inaccurate and noisy. To alleviate the problem, a number of approaches were proposed to output accurate disparity values for selected pixels only. Instead of designing another disparity optimization method for sparse disparity matching, we present a novel disparity filtering step that detects and removes inaccurate matches. Based on 3D convolutional neutral networks, our detector is trained directly on 3D matching cost volumes and hence work with different matching cost generation approaches. The experimental results show that it can effectively filter out mismatches while preserving the accurate ones. As a result, combining our approach with the simplest Winner-Take-All optimization will lead to a better performance than most existing sparse stereo matching algorithms on the Middlebury Stereo Evaluation site.