Efficient deep learning for stereo matching with larger image patches

Stereo matching plays an important role in many applications, such as Advanced Driver Assistance Systems, 3D reconstruction, navigation, etc. However it is still an open problem with many difficult. Most difficult are often occlusions, object boundaries, and low or repetitive textures. In this paper, we propose a method for processing the stereo matching problem. We propose an efficient convolutional neural network to measure how likely the two patches matched or not and use the similarity as their stereo matching cost. Then the cost is refined by stereo methods, such as semiglobal maching, subpixel interpolation, median filter, etc. Our architecture uses large image patches which makes the results more robust to texture-less or repetitive textures areas. We experiment our approach on the KITTI2015 dataset which obtain an error rate of 4.42% and only needs 0.8 second for each image pairs.