Attention aggregation learning for fast stereo matching

Abstract

In the stereo matching domain, deep learning has advanced significantly in extracting features from raw data. However, the complex networks yielding highly accurate disparity maps entail substantial computational expenses and longer processing time. Achieving a balance between real-time processing and accuracy poses a persistent challenge. Hence, we designed the Attention Bilateral Grid Network (ABGNet), a fast network leveraging the Attention Aggregation Module (AAM) and bilateral grid upsampling. For our network, initial contextual information is first obtained through lowresolution 3D convolution. Then, guided by a bilateral grid, edge-preserving up-sampling is applied to the coarse disparity map. Finally utilizing comprehensive contextual data, attention aggregation further refines the disparity map. This network ensures high-speed processing while maintaining specified accuracy levels.