Multi-Scale Context Attention Network for Stereo Matching

Recently, many works for stereo matching with convolutional neural networks have gained satisfactory performance. However, it is still an urgent challenge to deal with ill-posed regions and improve details in disparity maps. To address these problems, we propose a multi-scale context attention network with three main modules: atrous spatial pyramid pooling attention, richer convolutional features, and attention mechanism. First, we propose an atrous spatial pyramid pooling attention module to capture context information by the aggregating context in different scales, meanwhile take advantage of the attention mechanism to selectively emphasize informative features and suppress fewer ones. Then, the richer convolutional module is proposed to bring useful detail information for the network. Additionally, attention mechanism is used to pick out informative features for disparity refinement sub-network. Furthermore, we design a point-specific loss function strategy to perform online hard point mining, which helps the network to improve the accuracy of disparity maps. The experiments on the FlyingThings3D and KITTI 2015 benchmark demonstrate that the proposed method can achieve state-of-the-art performance.