Deep Census Weight Correlation Stereo Network

Stereo depth estimation is of great importance in autonomous driving, unmanned system operation and other intelligent visual tasks. At present, large textureless and uneven illumination regions are still the difficult problems for stereo matching tasks. To promote the stereo matching network to effectively utilize the feature information of the uneven illumination regions and the large textureless regions, we proposed the census weight correlation cost volume construction method which can selectively construct the illumination feature informaiton into the cost volume. The census weight correlation volume is a differentiable transformation of the traditional census-based matching cost computation. Extensive experiments show that the proposed census weight correlation volume can effectively improved the performance in the large textureless and uneven illumination regions compared with the original correlation volume and our network has achieved excellent performance on Scene Flow datasets, KITTI 2012 datasets and KITTI 2015 datasets.