Stereo Matching Via a Sparsity Constraint for Autonomous Driving

Stereo matching has recently achieved remarkable performance in computer vision applications, particularly in autonomous driving. A large number of studies have shown that the performance of stereo matching can be improved by fully utilizing the feature information of stereo matching image pairs to construct the cost volume. This work proposes an efficient cost volume construction method with sparsity constraint using information-theoretic principles, namely SCNet. Specifically, the features of the image pairs are first subjected to cascading calculations and correlation calculations to obtain the initial concatenation volume and the initial correlation volume. Then, information-theoretic principles are used to handle the initial correlation volume to enhance the correlation with larger-values and suppress the correlation with lower-values. Finally, the handled correlation and concatenation volumes are fused to construct the cost volume for stereo matching. The experiments show that the proposed SCNet has achieved advanced performance in stereo matching on three public datasets: KITTI2012, KITTI2015, and Scene Flow.