Binocular Stereo Matching Based on Convolutional Neural Networks

For the binocular stereo matching of deep learning based on patches, the networks structure is vital for matching cost in stereo matching. The task of using a pair of stereo images to estimate depth information can be achieved by a convolutional neural network after being formatted as a supervised learning task. However, the current stereo matching neural networks have poor stereo matching results in ill-posed-regions. In order to solve this problem, Our proposed a deep learning architecture that constructs a cost volume through improving the relationship between groups. The network consists of a feature extraction module, a cross-form spatial pyramid module and a feature matching fusion module. The improved stereo matching network is trained and verified on the KITTI data. The experimental results show that the improved network has certain advantages in terms of accuracy and speed compared with the previous methods.