A feature extraction technique in stereo matching network

Feature extraction is the basic and vital part of stereo matching. Accurate and effective feature extraction is the precondition of correct disparity estimation. Large-scale pooling is widely used to extract high-level features in the stereo matching network, but a lot of semantics is discarded. Meanwhile, frank bilinear interpolation with rare high-level information causes error propagation in the matching cost calculating. We apply Feature Pyramid Net to extract the multi-scale features with more context information and restore resolution layer by layer. The approach is validated on KITTI and Scene Flow.