GA-Stereo: A Real-Time Stereo Network Based on the Gradient Flow Shunting Strategy and the Atrous Pyramid Network

While stereo matching methods have great progress in recent years, they still face several formidable challenges. These challenges mainly include how to fully utilize global context and depth semantic information for depth feature extraction while maintaining low computational cost, and how to accurately predict disparity in object details and ill-posed regions of stereo images. In this paper, we propose GA-Stereo, an end-to-end stereo network aiming at fast and more accurate disparity regression from rectified stereo images. Firstly, we propose GA-Net, a depth feature extraction network for stereo networks; it utilizes a gradient flow shunting strategy and aggregated residual transformation to calculate the matching cost of stereo images. Secondly, we propose an Atrous Pyramid Network, which implements convolution kernels adaptive sampling to assist the initial disparity map optimize, on the basis of more feature information and larger receptive field. Finally, we propose a collaborative supervised learning strategy. By using the collaborative supervised learning strategy during training stage, our GA-Stereo can not only calculate loss from the Ground Truth, but also calculate the loss from perceptual reconstruction loss and edge-loss smoothing loss to obtain better training results and more accurate predicted disparity map than other stereo networks. Our GA-Stereo achieve highly competitive results in matching accuracy and computational efficiency on stereo benchmarks, including SceneFlow, KITTI2012, and KITTI2015. This advancement represents a significant step forward in real-time stereo matching methods.