Disparity Estimation with Deep Feature Parallel Transformation

We propose a feature parallel transformation module to strength the matching ability in stereo. We obtain additional information from input images, enhance features from the feature extraction module and design a parallel cascading module to fuse the features, which have rich global information. In addition, our feature transformation modules can be applied to any existing stereo networks in a plug and play manner. Our simple yet effect module can perform outstandingly on stereo benchmarks.