Soft Warping Based Unsupervised Domain Adaptation for Stereo Matching

Abstract

Stereo matching is a practical method to estimate depth information and retrieve 3D world in robot perception and autonomous driving scenarios. With the development of convolution neural networks (CNNs), deep-learning based stereo matching algorithms have significantly improved the accuracy and dominated most of the online benchmarks. However, limited labels in real world, especially in challenging weather conditions, still hinder the technology from practical usage. In this paper, we propose a new unsupervised learning mechanism for stereo matching, utilizing adversarial iterative learning and novel soft warping loss to promote the effectiveness of the networks in unseen environments. The experiments transferring the stereo matching module from synthetic domain to real-world domain demonstrate the superiority of our proposed method. Extensive experiments in challenging weathers further prove that our method shows great practical potential in strait environments.