Scholarlink

<p>Aligning extrinsically calibrated view sets is essential for merging 3D reconstructions from different agents or localizing them within a large existing map of the environment. In such scenarios, we not only have to account for the 3D rotation and translation but also need to estimate the unknown scaling factor between the reconstructions. In this paper, we propose the first closed-form solvers for image-only data to the general problem, leveraging either 26 point or 9 affine correspondences (AC) to obtain the scale, 3D orientation, and translation. Considering that modern image-capturing tools like smartphones and mixed reality devices typically come with Inertial Measurement Units and return the gravity direction by default, we also propose a solver that requires just 2 ACs along with gravity measurements. The proposed methods have been rigorously tested on both synthetic data and extensive publicly available real-world datasets. The results demonstrate that our approach achieves state-of-the-art accuracy and permits robust estimation in real-time owing to small sample sizes. The code is available at <a href="https://github.com/gowanting/GRPS-Affine">https://github.com/gowanting/GRPS-Affine</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generalized Relative Pose and Scale from Affine Correspondences

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Wanting Xu,
  • Xinyue Zhang,
  • Marc Pollefeys,
  • Daniel Barath,
  • Laurent Kneip

摘要

Aligning extrinsically calibrated view sets is essential for merging 3D reconstructions from different agents or localizing them within a large existing map of the environment. In such scenarios, we not only have to account for the 3D rotation and translation but also need to estimate the unknown scaling factor between the reconstructions. In this paper, we propose the first closed-form solvers for image-only data to the general problem, leveraging either 26 point or 9 affine correspondences (AC) to obtain the scale, 3D orientation, and translation. Considering that modern image-capturing tools like smartphones and mixed reality devices typically come with Inertial Measurement Units and return the gravity direction by default, we also propose a solver that requires just 2 ACs along with gravity measurements. The proposed methods have been rigorously tested on both synthetic data and extensive publicly available real-world datasets. The results demonstrate that our approach achieves state-of-the-art accuracy and permits robust estimation in real-time owing to small sample sizes. The code is available at https://github.com/gowanting/GRPS-Affine.