Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incremental Rotation Averaging Revisited

Xiang Gao, Hainan Cui, Yangdong Liu, Shuhan Shen

TL;DR

This work addresses large-scale rotation averaging for structure-from-motion by introducing IRAv4, an Incremental Rotation Averaging method that builds a task-specific, CDS-based alignment reference in tandem with estimating absolute rotations. By incrementally selecting Next-Best Vertex and updating rotations within a CDS-driven subgraph, IRAv4 enhances local-to-global rotation alignment robustness and scalability on noisy, richly connected graphs. Across extensive experiments on the 1DSfM dataset, IRAv4 achieves state-of-the-art accuracy in absolute rotation estimation and demonstrates superior performance in reference construction compared to prior IRA variants and mainstream baselines. The approach strengthens global SfM pipelines by providing a reliable, scalable mechanism for reference-driven alignment in challenging, large-scale scenes.

Abstract

In order to further advance the accuracy and robustness of the incremental parameter estimation-based rotation averaging methods, in this paper, a new member of the Incremental Rotation Averaging (IRA) family is introduced, which is termed as IRAv4. As its most significant feature, a task-specific connected dominating set is extracted in IRAv4 to serve as a more reliable and accurate reference for rotation local-to-global alignment. This alignment reference is incrementally constructed, together with the absolute rotations of the vertices belong to it simultaneously estimated. Comprehensive evaluations are performed on the 1DSfM dataset, by which the effectiveness of both the reference construction method and the entire rotation averaging pipeline proposed in this paper is demonstrated.

Incremental Rotation Averaging Revisited

TL;DR

This work addresses large-scale rotation averaging for structure-from-motion by introducing IRAv4, an Incremental Rotation Averaging method that builds a task-specific, CDS-based alignment reference in tandem with estimating absolute rotations. By incrementally selecting Next-Best Vertex and updating rotations within a CDS-driven subgraph, IRAv4 enhances local-to-global rotation alignment robustness and scalability on noisy, richly connected graphs. Across extensive experiments on the 1DSfM dataset, IRAv4 achieves state-of-the-art accuracy in absolute rotation estimation and demonstrates superior performance in reference construction compared to prior IRA variants and mainstream baselines. The approach strengthens global SfM pipelines by providing a reliable, scalable mechanism for reference-driven alignment in challenging, large-scale scenes.

Abstract

In order to further advance the accuracy and robustness of the incremental parameter estimation-based rotation averaging methods, in this paper, a new member of the Incremental Rotation Averaging (IRA) family is introduced, which is termed as IRAv4. As its most significant feature, a task-specific connected dominating set is extracted in IRAv4 to serve as a more reliable and accurate reference for rotation local-to-global alignment. This alignment reference is incrementally constructed, together with the absolute rotations of the vertices belong to it simultaneously estimated. Comprehensive evaluations are performed on the 1DSfM dataset, by which the effectiveness of both the reference construction method and the entire rotation averaging pipeline proposed in this paper is demonstrated.
Paper Structure (16 sections, 12 equations, 3 figures, 4 tables)

This paper contains 16 sections, 12 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Robust loss-based rotation averaging
  4. Outlier filtering-based rotation averaging
  5. Deep learning-based rotation averaging
  6. Brief description on the previous IRA series
  7. The proposed IRAv4 method
  8. Task-specific alignment reference construction
  9. Other key steps stemming from the IRAv3+
  10. Experimental evaluation
  11. Dataset and metric for evaluation
  12. Comparison in absolute rotation estimation
  13. Comparison results on synthetic data
  14. Comparison results on real-world data
  15. Comparison in alignment reference construction
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Toy examples of the traditional method (a) based on an approximation algorithm Guha-Algorithmica-98 and the task-specific method (b) proposed in IRAv4 for Connected Dominating Set (CDS) extraction. For (a), the red edges denote those between the currently selected vertex ($v^*$ or $v_i^\star$, where $v_i^\star$ denotes the selected vertex in the $i$-th iteration) and its adjacent unselected ones. And for (b), the red edges denote those between the selected triplet $t_{i,j,k}^*$ and the edge supporting set for the selected Next-Best Vertex (NBV) $v_{p^*}^i$ in the initialization step and the $i$-th iteration step, respectively. It could be observed from the figure that the number of vertices in the task-specific CDS is usually larger than that of the CDS extracted in the traditional way (5 vs. 3 in this toy example). Please refer to the main text for more details.
  • Figure 2: Synthetic experimental comparison results in absolute rotation estimation with the graph structures of MND and TOL, where the noise levels of $\sigma$ are set to $5\degree$ and $10\degree$. The $x$-axis and $y$-axis are the rotation median errors and edge outlier percentages.
  • Figure 3: Sparse scene reconstruction results on GDM by $4$ comparative rotation averaging methods, including IRLS-$\ell_{\frac{1}{2}}$Chatterjee-TPAMI-18, HARA Lee-CVPR-22, NeuRoRA Purkait-ECCV-20, and IRAv4, combined with translation averaging (BATA Zhuang-CVPR-18), multi-view triangulation Schonberger-CVPR-16, and global bundle adjustment Agarwal-ECCV-10.