Guided Cost Aggregation and Cost Distribution Refinement for EfficientNetV2-Based Stereo Matching Network

Stereo matching based on end-to-end deep learning architecture has been extensively studied for establishing stereo vision due to its high efficiency in the past few years. Accurate feature representation is crucial for subsequent steps in stereo matching. However, many existing stereo matching networks face challenges in extracting accurate features, which restricts the ability of generating excellent disparity maps. In this paper, we propose optimizing the feature extraction module based on EfficientNetV2 to enhance feature representation and establish a solid foundation for subsequent modules. In addition, current stereo matching networks struggle to effectively extract structural information from the cost volume during the cost aggregation stage. We introduce GCE to utilize feature information for generating guidance in order to improve the cost aggregation. In existing disparity refinement modules, the presence of static convolution kernels often results in over-smooth in disparity maps. Therefore, we propose to incorporate the cost self-reassembling (CSR) refinement strategy to optimize the cost distribution, leveraging its unimodal distribution property to mitigate the over-smoothing problem. Experimental results on the SceneFlow, KITTI 2012, and KITTI 2015 datasets demonstrate the effectiveness and superiority of our method. The source code is available at https://github.com/yuankang1234/EGR-Stereo.