Stereo matching on images based on volume fusion and disparity space attention

Constructing cost volumes by leveraging the feature information from image pairs is crucial to improve the stereo matching accuracy, which plays a vital role in acquiring depth information in autonomous driving system. In this work, we propose a convolutional neural network-based method called volume fusion and disparity space attention-based stereo matching model. The model mainly comprises two components: the volume fusion module and the disparity space attention module. Specifically, the volume fusion module initially computes a correlation volume and a concatenation volume. it then generates the initial cost volume based on these computed volume and concatenation volume. Finally, the disparity space attention module regularizes the features of the initial cost volume in two dimensions (i.e., the disparity dimension and the space dimension) to obtain the attention cost volume for stereo matching. Experimental results demonstrate that the proposed model outperforms several state-of-the-art models on three popular and public datasets.