Area-based correlation and non-local attention network for stereo matching

Stereo matching plays an essential role in various computer vision applications. Cost volume is the crucial part in disparity estimation for measuring the similarity between the left-right feature locations. However, most previous cost volume construction based on concatenation or pixel-wise correlation lack of local similarity, leads to an unsatisfactory performance on the large textureless regions. We propose a simple but efficient method for stereo matching to tackle the problem, called area-based correlation and non-local attention network (Abc-Net). First, we exploit the area-based correlation to capture more local similarity in cost volume. The left-right features are sliced into various size patches along the channel dimension. Correlation maps are calculated between the left feature patches and corresponding traversed right patches and then pack them into a 4D area-based cost volume. Second, based on the hourglass module, we combined it with the non-local attention module as the 3D feature matching module, which exploits various spatial relationships and global information. The experiments show that (1) the area-based correlation can capture local similarity to increase accuracy on the large textureless region, (2) the improved 3D feature matching module can exploit global context information to further improve performance, (3) our method achieves competitive results on the SceneFlow, KITTI 2012, and KITTI 2015 datasets.