Mask Reconstruction Augmentation and Attention Aggregation for Stereo Matching

Stereo matching is a significant task in computer vision and is widely used for depth prediction. Although deep learning stereo matching networks have achieved impressive results in recent years, the current models are prone to overfitting on synthetic datasets, resulting in poor generalization performance, and the ability of the current cost aggregation modules to perceive global context information is limited. In this paper, we propose a stereo matching framework with mask reconstruction augmentation and attention aggregation (MRA-AA) to address the above issues. Firstly, inspired by the work of applying masked image reconstruction to self-supervised learning, we design a novel data augmentation module suitable for stereo matching to improve the generalization performance of the model, namely mask reconstruction augmentation. Secondly, an attention aggregation module is proposed. The aggregation module can explicitly model the interdependencies between the channels of cost volume and capture the long-range dependence. Therefore, the module can help the model obtain more accurate disparity estimation in thin structures and texture-less regions. Furthermore, the convex upsampling module is used to upsample the cost volume to alleviate the blurred edges issue. Experimental results on Scene Flow and KITTI benchmarks show that our proposed method outperforms some existing methods.