Occlusion Aware Stereo Matching via Cooperative Unsupervised Learning

Abstract

Occlusion as a core challenge for stereo computation has attracted extensive research efforts in the past decades. Apart from its adverse impact, occlusion itself is a crucial clue which has not been exploited in the field of CNN based stereo. In this paper, we argue that a deep stereo framework benefits from reasoning occlusion in advance. We present an occlusion aware stereo network comprising a prior occlusion inferring module and a subsequent disparity computation module. The occlusion inferring module is a sub-network that directly starts from images, which averts the sophisticated procedure to iteratively estimate occlusion with disparity. We additionally propose cooperative unsupervised learning of occlusion and disparity, based on a different hybrid loss enforcing them to be consensus and trained alternatively to reach convergence. The comprehensive experimental analyses show that our method achieves state-of-the-art results among unsupervised learning frameworks, and is even comparable to several supervised methods.