结构化约束增强的6D物体位姿估计  

6D Object Pose Estimation Enhanced by Structural Constraint

作　　者：王立春 杨超[1,2] 付芳玉 WANG Lichun;YANG Chao;FU Fangyu(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Multimedia and Intelligent Software Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学多媒体与智能软件技术北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2025年第2期173-182,共10页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(62376014);中国高校产学研创新基金资助项目(2021JQR023)。

摘　　要：针对基于投票策略的6D物体位姿估计方法忽略了关键点间结构信息的问题,提出结构化约束增强的6D物体位姿估计方法——SC-Pose。该方法定义了一种用于描述物体2D关键点间结构化信息的形状描述符,通过增加关键点结构化损失约束形状描述符的预测值与真值相近,从而使2D关键点的定位更加准确,提升了6D物体位姿估计的精度。在LINEMOD、OCC-LINEMOD和TruncationLINEMOD数据集上进行了实验,结果表明,SC-Pose可以明显提升6D物体位姿估计的性能。Aiming at the problem that the 6D object pose estimation method based on the voting strategy ignores the structural information between keypoints,a 6D object pose estimation method enhanced by structural constraint,SC-Pose,is proposed.This method defines a shape descriptor to describe the structural information between the 2D keypoints of the object.By increasing the keypoint structural loss to constrain the predicted shape descriptor to be close to the ground-truth shape descriptor,the positioning of the 2D keypoints is more accurate,thereby ultimately enhancing the accuracy of 6D object pose estimation.Results on the LINEMOD,OCC-LINEMOD and TruncationLINEMOD datasets show that SC-Pose can significantly boost the accuracy of 6D object pose estimation.

关 键 词：6D物体位姿估计 单位向量场 投票策略 结构化损失 抓取交互 深度网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]