Multi-size Pooling for Stereo Matching Cost

Dense stereo matching is always a fundamental problem in computer vision due to its wide application related to three dimension scene reconstruction. In this paper, we present a framework to produce better unary features for stereo matching. To aggregate context information and take advantages of different representations, we first extract several features maps from multi-size patches with parameter-shared convolutional layers. Then we proposed two multi-size pooling modules to fuse spatial information. Experimental results prove that our proposed framework with both two pooling models improves the quality of unary features and effectively lower the error rates of disparity maps calculated from unary matching costs on both KITTI 2012 and KITTI 2015 datasets. Moreover, our proposed pooling modules can be easily applied to the feature extraction stage of end-to-end stereo matching frameworks.