Stereo matching based on attention mechanism and separable convolution

To improve the feature extraction capability of current stereo matching network and reduce the parameter, the multi-scale context attention network for stereo matching network is proposed. In the feature extraction stage, the improvedchannel-wise attention mechanism is used to weight the features based on the information contained in the channel, and the improved spatial pyramid structure is used to achieve multi-scale feature extraction to improve the network's feature extraction ability.A three-dimensional attention module and a three-dimensional separable convolution for disparity calculation.Compared with the standard convolution, the computational parameters of the network can be reduced and the channel dimension can be increased to ensure the matching accuracy. Experiments on Scene Flow datasets, KITTI 2012 datasets and KITTI 2015 datasets show that the network model in this paper can effectively reduce the calculation parameters of the network while ensuring matching accuracy.