Title: IRPSM-net: Information retention pyramid stereo matching network

Authors: Yun Zhao; Jiahui Tang; Xing Xu; Xiang Zhou

Addresses: School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Hangzhou, 310023, China ' School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Hangzhou, 310023, China ' School of Mechanical and Energy Engineering, Zhejiang University of Science and Technology, Hangzhou, 310023, China ' School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Hangzhou, 310023, China

Abstract: In order to prevent the lack of information in the stereo matching process and improve the disparity map accuracy. The information retention pyramid stereo matching network (IRPSM-Net) was proposed a novel architecture that can relieve the limitation of accuracy and retention the original information of the image. The proposed network consisted an information retention pyramid module (IRPM) without batch normalisation to retain the image information. And the training process was optimised by group normalisation, which further improves the effect of stereo matching. The ablation experiments show that our method can effectively improve the accuracy of 0.17% in the threshold 3 pixels of KITTI2012 stereo dataset and 0.09% in the whole region of KITTI2015 stereo dataset. It showed that the improvement of IRPSM-Net can effectively improve the quality of the generated disparity map.

Keywords: stereo matching; multi-scale; information retention pyramid; group normalisation.

DOI: 10.1504/IJCSM.2023.130426

International Journal of Computing Science and Mathematics, 2023 Vol.17 No.1, pp.50 - 66

Received: 02 Nov 2020
Accepted: 23 Jul 2021
Published online: 20 Apr 2023 *