«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年01期

页码:

53-60

栏目:

媒体计算

出版日期:

2025-01-10

文章信息/Info

Title:

A Method for 6D Object Pose and Size Estimation Based on Shape Priors

文章编号:

1673-629X(2025)01-0053-08

作者:

庞宇恒*; 贾迪

辽宁工程技术大学 电子与信息工程学院,辽宁 葫芦岛 125105

Author(s):

PANG Yu-heng*; JIA Di

School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China

关键词:

鲁棒性; 目标尺寸与6D姿态估计; 级联关系网络; 形状先验; 先验信息融合方法; 递归重构网络

Keywords:

robustness; 6D object size and pose estimation; cascaded relational network; shape prior; prior information fusion; recursive re-construction network

分类号:

TP391.41

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0258

摘要:

近年来,类别级 6D 目标尺寸与姿态估计逐渐成为研究热点,现有方法在处理复杂场景时鲁棒性不高,为此提出一种基于形状先验的目标尺寸与 6D 姿态估计方法。 首先,提出了一种级联关系网络来捕获多源输入的底层关系。 学习目标形状先验信息,优化网络结构增强形状特征表示能力。 其次,采用先验信息融合方法在特征融合前将形状先验特征进行融合,减小特征融合后的形变量,保留先验形状信息。 最后,设计了一个递归重构网络,将变形场输出与形状先验特征重新整合,提高模型对先验形状信息的保留能力。 与其他相关方法相比,该方法在 CAMERA25 数据集上(5?2 cm 姿态估计条件下)获得 95. 6% 的平均最优精度,在 REAL275 数据集的多个指标上较最优方法提升 5% ~ 10% ,消融实验验证了提出的三个模块的有效性。 该方法可以为后续高级应用提供更好的目标尺寸与 6D 姿态估计结果。

Abstract:

In recent years,category-level 6D object size and pose estimation have gradually become a research hotspot. Existing methods lack robustness in handling complex scenes. To address this,we propose a shape prior - based method for 6D object size and pose estimation. Firstly,a cascaded relational network is proposed to capture the underlying relationships of multiple source inputs. The target shape a priori information is learned and the network structure is optimized to enhance the shape feature representation capability.Secondly,a prior information fusion method is used to fuse shape a priori features before feature fusion,reduce the shape variable after feature fusion, and retain the a priori shape information. Finally, a recursive reconstruction network is designed to reintegrate the deformation field output with the shape a priori features to improve the model’s ability to retain the a priori shape information. Compared to other related methods,the proposed approach achieves 95. 6% average optimal precision on the CAMERA25 dataset under 5?2 cm pose estimation conditions and improves by 5% ~ 10% on multiple metrics on the REAL275 dataset compared to the best - performing methods. Ablation experiments validate the effectiveness of the three modules proposed. The proposed method can provide better results for 6D object size and pose estimation for subsequent advanced applications.

常用功能

更新日期/Last Update: 2025-01-10