Parallax Estimation Based on Transformer

In recent years, stereo matching employing neural networks has emerged as a crucial research direction within the domain of computer vision. Stereoscopic depth estimation hinges on the optimal correspondence between the polar pixels in the left and right images to infer the depth. To address the issue that the accuracy of the line-of-sight estimation task is prone to be disrupted by diverse factors in an unconstrained environment, this paper refines a feature cross-fusion approach based on convolution and the attention mechanism to accomplish parallax estimation. This method enhances the overall performance of the network by augmenting the efficacy of feature fusion, thereby improving the accuracy of line-of-sight estimation and imposing distinct constraints during matching.