Deep spatial and discriminative feature enhancement network for stereo matching

Stereo matching is of great importance to the environment perception, scene understanding, 3D reconstruction, autonomous driving and other tasks. Fully extracting the feature information of the input images and increasing the aggregation ability of the cost aggregation network are critical for stereo matching networks. To address this issue, we proposed a spatial feature enhancing and discriminative feature enhancing-based stereo matching network. For feature extraction, we proposed the spatial feature enhancing-based feature extraction network, which can fully extract the spatial position information and texture information of the input images. For cost aggregation, we proposed the discriminative feature enhancing-based cost aggregation network, which can enhance the selectively discriminative ability for useful feature information. Extensive experiment results show that our proposed method has achieved the state-of-the-art accuracy on Scene Flow datasets, KITTI2012 datasets, and KITTI2015 datasets.