华中科技大学学报（自然科学版）2025，Vol.53Issue(5)：78-84,7.DOI:10.13245/j.hust.250193
基于MobileGStereo的低复杂度立体匹配算法
Low complexity stereo matching algorithm based on MobileGStereo
摘要
Abstract
Aiming at the problem that the present advanced stereo matching models continuously improve the in-domain training accuracy by stacking deep modules,leading to an increase in computational cost and making it difficult to address the model degradation for cross-domain data,the sub-processes of stereo matching was reconstructed,such as feature extraction,cost computation and aggregation,and disparity refinement.By leveraging traditional models to make up for the deficiencies of deep neural networks,a low-complexity stereo matching model named MobileGStereo was proposed,which took into account both the generalization ability for cross-domain data and the fast inference ability.In the feature extraction stage,the deep network was designed to extract the differential features between pixels instead of complex semantic features.Meanwhile,the feature map should focus its representation distribution on itself rather than the entire batch.Therefore,MobileNet based on layer normalization was adopted as the backbone for feature extraction.In the cost computation and aggregation stage,a skip cost volume was proposed to reduce the computational complexity of high-resolution features in cost aggregation.By emulating the traditional aggregation method,a cross-scale aggregation method based on 3D depth-wise separable convolution was proposed to aggregate the cost volumes computed from features of different scales.Finally,a lightweight hourglass-like structure was utilized to fuse the multi-dimensional information of the cost after cross-scale aggregation and to regress the initial disparity.A ConvGRU-based structure was used to recurrently refine the initial disparity with the help of feature information.The proposed method was evaluated on the benchmark datasets.Experimental results show that the proposed method takes only 75 ms to predict about 1 226×370 resolution stereo images,which significantly improves the inference speed and achieves comparable quantitative performance with the state-of-the-art methods in the cross-domain data generalization ability test.关键词
深度学习/立体匹配/特征提取/代价计算与聚合/视差细化Key words
deep learning/stereo matching/feature extraction/cost computation and aggregation/disparity refinement分类
信息技术与安全科学引用本文复制引用
伍云霞,邹正阳,徐倩..基于MobileGStereo的低复杂度立体匹配算法[J].华中科技大学学报（自然科学版）,2025,53(5):78-84,7.基金项目
国家自然科学基金资助项目(52374165). （52374165）
