Points and Patches in Textures by Exploring in Monocular Depth Estimation

Rich texture information is sometimes not good for deep learning, and can introduce unwanted noise information, especially for monocular depth learning. In order to solve the phenomenon of ‘messy textures’ caused by rich texture in monocular depth estimation, this paper exploits the property of intra-patch disparity consistency to reconstruct the image to mitigate this phenomenon. The joint training method using patch matching and single-point matching is verified on Cityscapes and KITTI Engie Split datasets to achieve better results.