东北大学学报(自然科学版) ›› 2023, Vol. 44 ›› Issue (4): 457-468.DOI: 10.12068/j.issn.1005-3026.2023.04.001

• 信息与控制 •    下一篇

自适应构造与聚合多尺度代价体的双目立体匹配

庞彦伟, 苏畅, 龙涛   

  1. (天津大学 电气自动化与信息工程学院, 天津300072)
  • 发布日期:2023-04-27
  • 通讯作者: 庞彦伟
  • 作者简介:庞彦伟(1976-),男,河北衡水人,天津大学教授,博士生导师.
  • 基金资助:
    科技创新2030-“新一代人工智能”重大项目(2022ZD0160400); 天津市科技计划项目(19ZXZNGX00050).

Adaptive Multi-scale Cost Volume Construction and Aggregation for Stereo Matching

PANG Yan-wei, SU Chang, LONG Tao   

  1. School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China.
  • Published:2023-04-27
  • Contact: PANG Yan-wei
  • About author:-
  • Supported by:
    -

摘要: 基于卷积神经网络的双目立体匹配算法取得了重要进展,但现有方法在弱纹理区域、细节和边缘等位置仍然存在匹配不准确的问题.立足于双目立体匹配任务中常用的匹配代价体(cost volume),提出自适应构造与聚合多尺度代价体的双目立体匹配网络.首先将多个尺度的输入特征融合成为重组特征;然后设计可学习的特征增强模块,为各个尺度的匹配代价体恢复所需的细节信息;最后基于全局注意力对各尺度匹配代价体进行尺度内聚合,并提出自适应多尺度加权方法进行尺度间聚合,筛选出适用于回归各尺度视差的匹配特征.在SceneFlow和KITTI2015数据集上的实验表明:所提方法在较小网络规模的情况下取得了有竞争力的性能表现,验证了所提方法的有效性.

关键词: 双目立体匹配;匹配代价体;特征增强;自适应聚合

Abstract: Stereo matching based on convolutional neural network has made great progress. Existing methods still suffer from mismatching in weak texture regions, details and edges. Based on the cost volume commonly used in stereo matching, a stereo matching network with adaptive multi-scale cost volume construction and aggregation was proposed. Firstly, the proposed method fully fused the multi-scale features to obtain the recombined features. Then, a learnable feature enhancement module was used to recover the detail information for multi-scale cost volumes. Finally, after intra-scale aggregation based on global attention, an adaptive multi-scale weighting method was proposed for inter-scale aggregation to screen the matching features adapted to the disparity regression of each scale. Massive experiments on the SceneFlow and KITTI2015 datasets show that the proposed method achieves competitive performance with smaller network size which verifies the effectiveness of the proposed method.

Key words: stereo matching; cost volume; feature enhancement; adaptive aggregation

中图分类号: