Multiple Frequency Inputs and Context-Guided Attention Network for Stereo Disparity Estimation

Deep learning-based methods have made remarkable progress for stereo matching in terms of accuracy. However, two issues still hinder producing a perfect disparity map: (1) blurred boundaries and the discontinuous continuous region on estimation maps, (2) lack effective means to restore resolution precisely. In this paper, we propose utilize multiple frequency inputs an attention mechanism construct deep model. Specifically, high-frequency low-frequency information input image together with RGB are fed into feature extraction network 2D convolutions. It is conducive produce distinct boundary smooth maps. To regularize 4D cost volume regression, 3D context-guided module stacked hourglass networks, where high-level volumes as context guide low-level features obtain high-resolution yet precise The proposed approach achieves competitive performance SceneFlow KITTI 2015 datasets.