Liu, Changlin; Zhang, Yuan; Sun, Linjun; Zhang, Kaijie; Cao, Bing; Li, Zhiwei Source: 2023 International Conference on High Performance Big Data and Intelligent Systems, HDIS 2023, p 131-137, 2023, 2023 International Conference on High Performance Big Data and Intelligent Systems, HDIS 2023;
Abstract:
In stereo matching, perspective differences in images can cause feature inconsistencies. Traditional stereo matching algorithms compare pixel disparities using local parallel windows transformed by matching costs. However, this method under the fronto-parallel windows assumption may produce incorrect matching results when there are perspective differences in the feature scale. To address this issue, we propose a new cost computation method called relative gradient matching cost (RGC). RGC effectively handles feature perspective differences by robustly constraining feature matching relationships across varying object surface depths, simulating the effect of slanted support windows. We propose different improvement strategies for common traditional methods and DNN-based methods combined with RGC. In experiments, we compare the performance of RGC with some SOTA models. The experiment results demonstrate that RGC significantly improves the accuracy and stability of stereo matching, especially in cases with edge and discontinuity areas.
©2023 IEEE. (33 refs.)