Unsupervised Learning for Stereo Depth Estimation using Efficient Correspondence Matching

To avoid using costly ground-truth, several recent methods formulate the unsupervised depth estimation problem as the task of disparity estimation. However, most of them for monocular depth estimation cannot perform explicit depth maps, while even some methods for stereo depth estimation cannot efficiently use stereo-view information. In this work, we provide a novel cost volume construction approach (matching cost layer) to extract easy-to-learn cost volume by matching feature correspondences between stereo images, which can satisfy the unsupervised learning strategy better. We also propose a new network architecture that learns more reliable spatial location information of objects from fully fused stereo features. Extensive experiments of our method on the KITTI datasets demonstrate its superiority over the state-of-the-art methods.