Unsupervised detail-preserving network for high quality monocular depth estimation

In this paper, we propose an unsupervised learning framework to address the problems of the inaccurate inference of depth details and the loss of spatial information for monocular depth estimation. First, as an unsupervised technique, the proposed framework takes easily collected stereo image pairs instead of ground truth depth data as inputs for training. Second, we design a rectangle convolution to capture global dependencies between neighboring pixels across entire rows or columns in an image, which can bring significant promotion on depth details inference. Third, we propose a learned depth refinement module including a color-guided refinement layer and a learned composite proximal operator to preserve depth discontinuities and obtain high quality depth map. The proposed network is fully differentiable and end-to-end trainable. Extensive experiments evaluated on KITTI, Cityscapes and Make3D dataset demonstrate our state-of-the-art performance and good cross-dataset generalization ability.