Unsupervised Hierarchical Iterative Tile Refinement Network With 3D Planar Segmentation Loss

Abstract

Unsupervised real-time stereo matching is of great research value in robot navigation due to its independence from ground truth and real-time efficiency. The core challenge lies in the design of loss functions that can provide accurate guidance and the efficient network architectures. The commonly used photometric loss is prone to provide incorrect guidance because of the influence of reflection, left-right color inconsistency, low texture, and occlusion. As a weak supplement, the smoothness loss can ameliorate the multi-solution problems caused by low texture, but it is not effective for strong incorrect guidance caused by the other problems. In order to provide more accurate and powerful supplementary guidance, a 3D planar segmentation loss is proposed with advancements in addressing the strong incorrect guidance problem, which could be generally integrated into traditional unsupervised training losses. Furthermore, the real-time stereo matching approach of the hierarchical iterative tile refinement network is applied to unsupervised stereo matching, with necessary modifications to address the detrimental architectures that hinder its performance in unsupervised training. Experimental results verify the effectiveness of the 3D planar segmentation loss and the network modification. The proposed pipeline achieves competitive accuracy compared to existing unsupervised stereo matching methods while maintaining real-time efficiency.