Hierarchical Correlation Stereo Matching Network

Recently, stereo matching from a pair of rectified images has been cast as a supervised learning task using the powerful representation of convolutional neural networks. However, existing methods only utilize last feature maps output from Siamese Networks to compute similarity measurement, which are lack of multi-levels similarity information to construct an informative cost volume. To solve this problem, we propose a hierarchical correlation operation to compute similarity of stereo pairs at multiple levels. In addition, to yield accurate disparity in ill-posed region, we propose a stacked hourglass feature network with dense connections to effectively incorporate context information. Then, hybrid matching cost volume is built leveraging hierarchical correlation features and concatenation features of left and right. 3D CNN encoder-decoder architecture is utilized to regularize the cost volume and regress disparity. Experiments demonstrate that our network achieves competitive performance with state-of-the-art methods on Scene Flow, KITTI 2012, and KITTI 2015 datasets.