Uncertainty Estimation with Evidential Regression for Lightweight Stereo Matching

Stereo matching is a significant part in many computer vision tasks, particularly in autonomous driving and advanced driver assistance systems. Even though existing stereo matching methods have achieved considerable progress, they often exhibit pronounced disparity jumps at edges and thin structures, and lack uncertainty estimation for the predicted disparity maps. These issues are crucial for ensuring the safe operation of vehicles. In this paper, we introduce a lightweight stereo matching architecture, which leverages deep evidence regression to simultaneously estimate disparity maps and their associated uncertainties. During this process, we effectively integrate contextual features, enhancing the performance on edge details. To improve real-time performance, we propose the Group Multi-axis Attention Volume (GMAV), a concise cost volume that uses correlation volumes as attention weights to refine feature volume processed by Group Multi-axis Attention Fusion (GMAF). Based on these two structures, we designed a stereo matching method called Uncertainty Estimation with Evidential Regression Network (UERNet). Experiments on SceneFlow and KITTI datasets demonstrate that our method achieves a good balance between speed and accuracy. Our network achieved a D1-all accuracy of 2.01% on the KITTI 2015 test set. This represents a 20% performance increase over the baseline model.