A hybrid stereo matching framework

Abstract

In nowadays, substantial advancements have been achieved in stereo matching through deep learning techniques. However, in practical applications, traditional models still suffer from substantial computational overhead, making it challenging to meet the demands of both efficiency and accuracy. In the dissertation, a novel stereo matching framework combining PSMNet and STTR is proposed to address the bottleneck. By leveraging the mechanism of global attention within the Transformer module of STTR, the framework achieves raw disparity estimation with high quality. Based on raw disparity guidance, a sliding window strategy is introduced to construct a cost volume filled with residual matching costs, significantly reducing memory consumption. Additionally, the integration of spatial pyramid pooling and residual network structure enhances the framework in multi-scale feature fusion. Experimental findings reveal that the model performs well on various datasets, considerably improving computational efficiency while maintaining high disparity accuracy.