Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (21): 182-188.DOI: 10.3778/j.issn.1002-8331.2104-0025

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Stereo Matching Network Based on Edge-Guided Feature Fusion and Cost Aggregation

ZHANG Haodong, SONG Jiafei, ZHANG Guanghui   

  1. 1.Bio-Vision System Laboratory, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China
    2.University of Chinese Academy of Sciences, Beijing 100049, China
    3.ShanghaiTech University, Shanghai 201210, China
  • Online:2022-11-01 Published:2022-11-01



  1. 1.中国科学院 上海微系统与信息技术研究所 仿生视觉系统实验室,上海 200050
    2.中国科学院大学,北京 100049
    3.上海科技大学,上海 201210

Abstract: Aiming at the problem of large errors in the fine structure of stereo matching, especially at the edges, a stereo matching algorithm using edge-guided feature fusion and cost aggregation is proposed. The edge of the image is used to guide the weighted fusion of features of different scales, that is, the edges of small-scale features and the non-edges of large-scale features are given more weight to obtain fused features with stronger characterization ability. In the cost aggregation stage, the matching cost at the edge is weakened and unreliable information dissemination is reduced. The proposed method is evaluated on the SceneFlow and KITTI 2015 datasets, and the errors of the benchmark network PSMNet are reduced by 35.2% and 2.2%, respectively. Experiments have proved that the introduction of edge information has specifically improved the disparity solution of existing algorithms at fine structures(especially at the edges), and improves the overall prediction accuracy. In addition, the mentioned module is lightweight and can be applied to different stereo matching networks.

Key words: machine vision, stereo matching, convolutional neural network, binocular vision, edge information

摘要: 针对立体匹配在精细结构,尤其边缘处的误差较大的问题,提出了利用边缘引导特征融合和代价聚合的立体匹配算法。利用图像边缘引导不同尺度特征体加权融合,即对小尺度特征体的边缘处,大尺度特征体的非边缘处赋予更大权重,以获得表征能力更强的融合特征体。在代价聚合阶段弱化边缘处匹配代价,减少不可靠信息传播。所提方法在SceneFlow和KITTI 2015数据集进行了评估,将基准网络PSMNet的误差分别降低了35.2%、2.2%。实验证明,边缘信息的引入针对性地改善了现有算法在精细结构处(尤其边缘处)的视差求解,提高了整体预测精度。此外,所提的模块是轻量的,可适用于不同的立体匹配网络。

关键词: 机器视觉, 立体匹配, 卷积神经网络, 双目视觉, 边缘信息