Depth Estimation with Multi-Resolution Stereo Matching

Depth estimation has widely demands in autopilot and scene reconstruction. Although depth estimation has been greatly improved by deep learning, there is still some room for improvement. Usually, stereo-matching-based depth estimation matches low-resolution features and then up-samples depth map to full-resolution. Such methods suffer from low accuracy because of information loss in low-resolution features. To solve this problem, a depth estimation method via multi-resolution gradual-refining stereo matching is proposed. As with classic methods, this method first extracts pyramid features by convolution network and estimates low-resolution depth maps. Then, as an innovation, stereo-matching at each pyramid feature is successively executed using the low-resolution maps as initial depth, which limits search range of stereo-matching for accuracy and efficiency. Results of stereo-matching are used as residuals to refine the depth maps. Experimental results demonstrate that accuracy of depth estimation by the proposed method is significantly improved, but computational complexity does not much increase compared with classic methods.