Adaptively identify and refine ill-posed regions for accurate stereo matching

Abstract

Stereo matching cost constrains the consistency between pixel pairs. However, the consistency constraint becomes unreliable in ill-posed regions such as occluded or ambiguous regions of the images, making it difficult to explore hidden correspondences. To address this challenge, we introduce an Error-area Feature Refinement Mechanism (EFR) that supplies context features for ill-posed regions. In EFR, we innovatively obtain the suspected error region according to aggregation perturbations, then a simple Transformer module is designed to synthesize global context and correspondence relation with the identified error mask. To better overcome existing texture overfitting, we put forward a Dual-constraint Cost Volume (DCV) that integrates supplementary constraints. This effectively improves the robustness and diversity of disparity clues, resulting in enhanced details and structural accuracy. Finally, we propose a highly accurate stereo matching network called Error-rectify Feature Guided Stereo Matching Network (ERCNet), which is based on DCV and EFR. We evaluate our model on several benchmark datasets, achieving state-of-the-art performance and demonstrating excellent generalization across datasets. The code is available at https://github.com/dean7liu/ERCNet_2023.