北京大学学报（自然科学版） ›› 2026, Vol. 62 ›› Issue (5): 999-1014.DOI: 10.13209/j.0479-8023.2026.104

上一篇    下一篇

图像三维重建下视差角与笛卡尔参数化的光束法平差收敛性对比——基于百余组数据集的实验验证

谭平1, 代云凡2,3, 左正康2,3,4,†, 高云龙4, 刘东霞3,5, 董娇娇3, 王嘉伟2,3, 朱桂海4, 徐二帅6, 王强7, 卢文艺8, 赵李杰9, 杨冲4, 李小凯4, 刘文轩4, 杨洋4, 韩秉宏10
  

  1. 1. 中国电建集团成都勘测设计研究院有限公司, 成都 610000 2. 极坐标视觉实验室, 太原理工大学, 太原 030024 3. X-DNA智能文化联合实验室, 太原理工大学, 太原 030024 4. 大势智慧科技有限公司研发中心, 武汉 430223 5. 太原理工大学建筑与艺术学院, 太原 030024 6. 河北农业大学城乡建设学院, 保定 071001 7. 天津师范大学地理学部, 天津 300387 8. 中国地质大学地理与信息工程学院, 武汉 430078 9. 山西中云智图地理信息科技有限公司, 太原 030006 10. 中铁建工集团有限公司, 北京 100160
  • 收稿日期:2025-07-10 修回日期:2026-05-14 出版日期:2026-09-20 发布日期:2026-09-20
  • 基金资助:
    国家自然科学基金青年科学基金项目(C 类) (T2600787)、山西省青年科学研究项目(202403021212106)和太原理工大学引进人才科研启动经费(RY2400000591)资助

Comparison of Bundle Adjustment Convergence Using Parallax-Angle and Cartesian Parameterizations for Image-Based 3D Reconstruction: Experimental Validation across over 100 Datasets

TAN Ping1, DAI Yunfan2,3, ZUO Zhengkang2,3,4,†, GAO Yunlong4, LIU Dongxia3,5, DONG Jiaojiao3, WANG Jiawei2,3, ZHU Guihai4, XU Ershuai6, WANG Qiang7, LU Wenyi8, ZHAO Lijie9, YANG Chong4, LI Xiaokai4, LIU Wenxuan4, YANG Yang4, HAN Binghong10
  

  1. 1. POWERCHINA Chengdu Engineering Corporation Limited, Chengdu 610000 2. Polar-vision Lab, Taiyuan University of Technology, Taiyuan 030024 3. X-DNA Intelligent Culture Joint Lab, Taiyuan University of Technology, Taiyuan 030024 4. R&D Center, Daspatial Technology Co., Ltd., Wuhan 430223 5. College of Architecture and Art, Taiyuan University of Technology, Taiyuan 030024 6. Department of Urban and Rural Construction, Hebei Agricultural University, Baoding 071001 7. Faculty of Geography, Tianjin Normal University, Tianjin 300387 8. School of Geography and Information Engineering, China University of Geosciences, Wuhan 430078 9. Shanxi Zhongyun Zhitu Geoinformation Technology Co., Ltd., Taiyuan 030006 10. China Railway Construction Engineering Group Co., Ltd, Beijing, 100160
  • Received:2025-07-10 Revised:2026-05-14 Online:2026-09-20 Published:2026-09-20

PDF

摘要：

针对视差角参数化方法用于图像三维重建时存在与主流方法的对比实验规模有限、优化器设置不统一的问题, 构建涵盖车载场景、无人机场景和近景场景的106组光束法平差(BA)数据集, 对主流的笛卡尔坐标参数化(SBA)和视差角参数化方法(PBA)的求解流程进行统一, 确保对比的公平性, 并从收敛性、计算耗时、图像空间精度和参数空间精度等维度对两类方法进行评估。实验结果表明, 在106组配对数据集上, PBA方法的迭代次数整体上显著低于SBA方法, 呈现更快的收敛趋势; 在初始点云噪声较大或 SBA方法出现发散的代表性场景中, PBA方法在图像空间和参数空间中的优化结果更稳定。本文的数据集与源码已在https://reurl.cc/OmZz09和https://github.com/Polar-vision/PBA-V3公布。

关键词: 视差角, 笛卡尔, 三维点参数化, BA数据集, 收敛性

Abstract:

To address the problems of limited-scale comparative experiments against mainstream approaches and inconsistent optimizer configurations when applying parallax‑angle parameterization to image‑based 3D reconstruction, 106 bundle adjustment (BA) datasets were constructed spanning vehicle‑borne, UAV, and close‑range scenarios. We standardized the solution workflows for the mainstream Cartesian-coordinate parameterization (SBA) and parallax-angle parameterization (PBA) methods to ensure a fair comparison. The two methods were systematically evaluated in terms of convergence, computational time, image-space accuracy, and parameter-space accuracy. The experimental results showed that, across the 106 paired datasets, PBA generally required significantly fewer iterations than SBA, exhibiting a faster convergence trend. In representative scenarios with relatively large initial point-cloud noise or divergence of the Cartesian-coordinate method, PBA yielded more stable optimization results in both image and parameter spaces. The datasets and source code are publicly available at https://reurl.cc/OmZz09 and https://github.com/Polar-vision/PBA-V3. 

Key words: parallax-angle, Cartesian, 3D point parameterization, BA dataset, convergence

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司