A Normalized Disparity Loss for Stereo Matching Networks

Abstract

Stereo vision is an essential component for lots of robot vision applications. The performance of stereo matching has been greatly improved recently thanks to deep learning technologies. However, most of the current deep networks directly use naive disparity loss to train the network, regardless of the inherent unbalance of its supervision to pixels with different depths. In this letter, we propose a normalized disparity loss to solve this problem. The new loss is constructed as the common disparity loss divided by a fitted cost function. This function is responsible for imitating the variance of the common loss along the disparity dimension. Range restriction and gradual updating strategies are further designed to ensure the robustness of the loss function during training. Experimental results on multiple public datasets show that our method can be directly applied to most state-of-the-art networks and effectively improve their disparity performance.