Adaptive Neighbor Embedding for Efficient Stereo Matching

With the remarkable achievements of stereo matching algorithms based on Convolutional neural Networks (CNNs), more and more stereo matching algorithms based on CNNs are applied to intelligent driving research. Whereas, the existing convolution operators cannot fully mine and utilize the structure information, which is indispensable for stereo matching, especially in the regions with discontinuous disparity. In this work, to ameliorate these issues, an adaptive neighbor embedding paradigm (ANE paradigm) is proposed for deep stereo networks. In our ANE paradigm, the discriminative ability of features is improved by mining the neighbor correlation knowledge and performing adaptive neighbor aggregation. Specifically, we design a new convolution operator, termed adaptive neighbor embedding convolution (ANE conv) and its simplified version, termed adaptive neighbor embedding filter (ANE filter). Both ANE conv and ANE filter can be used as plug-and-play operators for all existing stereo matching models and can be easily trained end-to-end by standard back-propagation. Extensive experiments emphasize the performance of our approaches, especially in high accuracy metric.