基于引导优化的立体匹配网络  

Guided refinement for stereo matching network

作　　者：李杰 昌明源 向泽林[2] 都双丽 梁敏[1] 李旭伟[4] LI Jie;CHANG Ming-Yuan;XIANG Ze-Lin;DU Shuang-Li;LIANG Min;LI Xu-Wei(School of Information,Shanxi University of Finance and Economics,Taiyuan 030006,China;Chengdu College,Sichuan International Studies University,Chengdu 611844,China;Computer Science and Engineering,Xi'an University of Technology,Xi'an 710048,China;School of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[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

出　　处：《四川大学学报（自然科学版）》2024年第4期232-243,共12页Journal of Sichuan University(Natural Science Edition)

基　　金：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),山西省基础研究计划青年科学研究项目(202103021223308);山西省高等学校哲学社会科学研究项目(2021W058);西安碑林区应用技术研发项目(GX2244)。

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.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 constructed,utilizing guided deformable convolution.Unlike deformable convolution,this module performs offset 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 designed based on the guided refinement module.This network introduces a three-level cascaded cost aggregation 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 module on the KITTI2015 datasets indicate that the D1-noc and D1-all values of advanced algorithms such as GwcNet 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 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]