Stereo matching network based on AANet+ improved attention mechanism
Jiawen Wang, Yong Hua Yin · 2021
The stereo matching method based on deep learning is developing rapidly. Before the AANet+ adaptive aggregate stereoscopic matching network was put forward, the 3D convolution-based deep learning network occupied the core position, but it had the disadvantage of large number of parameters and high memory consumption in cost volume and cost aggregation. With the advent of AANet+, it solves the defects of 3D convolution to some extent by virtue of deformable convolution and intra-scale aggregation modules and cross-scale aggregation modules. However, there are also shortcomings in AANet+, which uses hourglass stacking modules to start feature extraction from multiscale directions, but uses convolution to lose some of its information when fused. Therefore, this paper uses the channel attention module ECA to improve the feature extraction module of AANet+, and introduces the channel attention module in the traditional hourglass stacking module to extract more rich features. In addition, channel attention modules are introduced to obtain multi-scale features to reduce information loss and increase contextual connections. Experimental results show that the proposed improved network has achieved better results in synthetic data set Scene Flow and realworld scene data set KITTI2015 than AANet+.