Attention Aggregation Encoder-Decoder Network Framework for Stereo Matching

Abstract

In the stereo matching networks based on deep learning, current cost aggregation networks lack the means to aggregate cost volume to the utmost extent. Therefore, different from the standard encoder-decoder structures, we propose an attention aggregation encoder-decoder network framework for stereo matching that contains three modules. Specifically, we design a sub-branch and cross-stage aggregation encoding module, which aggregate context information of different sub-branches and cross-stages to achieve the mutual utilization of different deep cost volumes. Meanwhile, we introduce a three-dimensional attention recoding module to obtain the robust discriminative cost volume through recalibrating the high-level semantic information of the sub-branches. In addition, we construct a stepwise aggregation decoding module to decode the cost volume via the stepwise fusion upsampling strategy, which further enhances the learning ability of the network model. The experimental results on Scene Flow and KITTI benchmark datasets show that the proposed network framework is superior to other similar methods in aggregating information.