Graph Convolution Network for Superpixel and Stereo Matching

We propose a method to predict depth using stereo images. Existing methods often suffer from pixel mismatch in small targets and ill-posed regions. To address this problem, we use superpixel segmentation to group pixels with similar color information into larger units, and learn the weight information between adjacent superpixels through a graph convolutional network. We then use the learned superpixel information to guide the disparity map prediction of the stereo matching architecture. Furthermore, we introduce superpixel pooling and superpixel correlation cost volume to constrain the stereo matching features, providing an effective architecture for cost aggregation. By incorporating superpixel information, we improve the accuracy of existing stereo matching models. Experimental results demonstrate that the proposed method significantly improves the performance of the depth prediction.