AHSM-Net: Unsupervised Stereo Matching Algorithm Based on Attention Mechanism and Hybrid Dilated Convolution
Xiaoge Li, Xingfang Zhao, Long Yan · Journal of Physics Conference Series · 2022
Abstract Too many parameters often accompany the stereo matching algorithm of the convolutional neural network. The dense matching makes the constructed cost volume increase with the resolution, resulting in high occupied memory and processing time. To perform depth estimation more efficiently, this paper proposes a stereo matching algorithm AHSM-Net based on attention mechanism and hybrid dilated convolution (HDC), which uses an unsupervised method to train the stereo matching network end-to-end. This method obviously enhances the consistency of objects. AHSM-Net is trained on the SceneFlow dataset, and a large number of experiments and analyses are carried out on the KITTI2012 and KITTI2015 datasets, which verifies that the method proposed in this paper has great matching accuracy and speed under the condition of unsupervised.