Deep learning-based disparity-shift stereo matching network

Although the existing binocular stereo matching method based on 3D convolution can retain more 3D geometric information, its computational efficiency is not high. To solve the above problems, this paper improves Gwc-Net and proposes an efficient and high-performance disparity displacement stereo matching model. The feature extraction network of this model is based on deep separable convolution, which introduces learnable dilation coefficients and jump connections to improve adaptability to different data and computational efficiency. Secondly, this paper extends the displacement operation to the disparity domain for the first time and achieves the purpose of aggregating neighborhood disparity information through 2D convolution through displacement in the disparity dimension. In experiments, the method proposed in this paper has higher detection accuracy on the SceneFlow dataset.