A Multi-Correlation Stereo Matching Network for Disparity Estimation

Recent Convolutional Neural Networks(CNNs) based networks for disparity estimation has a better performance in both prediction accuracy and time efficiency compared with traditional hand-crafted feature based methods. However, to achieve SOTA accuracy, CNN models tend to require a large amount of memory and computing resources, especially for those networks with 3D convolution. Hence we propose a compact and effective 2D convolution based network for disparity estimation, called MCSMNet, which has three main features:1) It introduces the basic units of ShuffleNet[1] and SENet[2] to make the training of network more effective without reducing network parameters; 2)It contains Multi-Correlation so that the context information could be fused in high-level features well; 3)It uses hierarchical refinement strategy to further improve the accuracy of networks. We conduct experiments to demonstrate the effectiveness of MCSMNet on the popular datasets, Scene Flow. Experimental results show that MCSMNet achieves competitive accuracy compared with existing 3D convolution based networks.