基于引导优化的立体匹配网络
Guided refinement for stereo matching network
李杰 1昌明源 1向泽林 2都双丽 3梁敏 1李旭伟4
作者信息
- 1. 山西财经大学 信息学院,太原 030006
- 2. 四川外国语大学 成都学院,成都 611844
- 3. 西安理工大学 计算机科学与工程学院,西安 710048
- 4. 四川大学 计算机学院,成都 610065
- 折叠
摘要
为克服细节区域精细立体匹配问题,本文提出了基于引导优化的立体匹配网络.首先,构建基于引导可变形卷积的引导优化模块,不同于可变形卷积,该模块对额外输入的引导特征进行偏移量和调制标量学习,增强可变形卷积的变形参数学习能力.其次,设计基于引导优化模块的引导优化立体匹配网络,该网络提出了基于3D代价聚合和2D引导优化聚合的三级串联代价聚合模块,逐步优化细节区域的配准精度.实验结果显示,在SceneFlow、KITTI等标准数据集中,与先进算法相比,该算法可实现细节区域的高精度配准.其中,引导优化模块适用性测试结果显示,在KITTI2015数据集中,增加引导优化模块后GwcNet、AANet等先进算法的D1-noc、D1-all值均产生20%左右的提升.
Abstract
Numerous challenges exist in achieving high-precision stereo matching for intricate areas,such as small structures and edge regions.To address the issue of fine stereo matching in detailed areas,we propose a stereo matching network based on guided refinement is proposed.Firstly,a guided refinement module is con-structed,utilizing guided deformable convolution.Unlike deformable convolution,this module performs off-set and modulation scalar learning on additional input guide features to enhance the deformation parameter learning ability of deformable convolution.Secondly,a guided refinement stereo matching network is de-signed based on the guided refinement module.This network introduces a three-level cascaded cost aggrega-tion module,incorporating 3D cost aggregation and 2D guided refinement aggregation,progressively refining the registration accuracy of detailed region.Experimental results demonstrate that,compared with state-of-the-art algorithms on standard datasets such as SceneFlow and KITTI,the proposed algorithm achieves high-precision registration of detailed regions.Notably,the applicability test results of the guided refinement mod-ule on the KITTI2015 datasets indicate that the D1-noc and D1-all values of advanced algorithms such as Gw-cNet and AANet increase by approximately 20%after integrating the guided refinement module.
关键词
立体匹配/引导可变形卷积/引导聚合/多特征提取/边缘保持Key words
Stereo matching/Guided deformable convolution/Guide aggregation/Multi-feature extraction/Edge-preserving引用本文复制引用
基金项目
国家自然科学基金项目(61801279)
山西省基础研究计划自然科学研究项目(202203021211333)
山西省高等学校哲学社会科学研究项目(2021W058)
山西省基础研究计划青年科学研究项目(202103021223308)
西安碑林区应用技术研发项目(GX2244)
出版年
2024
国家科技期刊平台
NSTL
万方数据