Linear Relative Pose Estimation Founded on Pose-only Imaging Geometry

TL;DR

The paper tackles robust two-view relative pose estimation in the presence of substantial outliers and degeneracies by introducing LiRP, a linear relative pose method grounded in pose-only imaging geometry. By integrating LiGT-based constraints into IRLS and RANSAC, the approach screens outliers and remains effective in planar or pure-rotation scenarios. A six-point LiRP algorithm yields multiple candidate solutions, which are disambiguated using pose-only residuals, and the robustness is enhanced with a GNC-IRLS/RANSAC fusion. Experimental results on synthetic data and the Strecha dataset show up to $2$-$10\times$ improvements in rotation accuracy under outlier fractions as high as $80\%$, highlighting practical impact for SLAM/SfM pipelines, especially in challenging textures and degenerate scenes.

Abstract

How to efficiently and accurately handle image matching outliers is a critical issue in two-view relative estimation. The prevailing RANSAC method necessitates that the minimal point pairs be inliers. This paper introduces a linear relative pose estimation algorithm for n $( n \geq 6$) point pairs, which is founded on the recent pose-only imaging geometry to filter out outliers by proper reweighting. The proposed algorithm is able to handle planar degenerate scenes, and enhance robustness and accuracy in the presence of a substantial ratio of outliers. Specifically, we embed the linear global translation (LiGT) constraint into the strategies of iteratively reweighted least-squares (IRLS) and RANSAC so as to realize robust outlier removal. Simulations and real tests of the Strecha dataset show that the proposed algorithm achieves relative rotation accuracy improvement of 2 $\sim$ 10 times in face of as large as 80% outliers.

Figures (8)

• Figure 1: Simulation of two-view structure and 3D feature points. The green nodes represent 3D points that conform to chirality constraints, and the black nodes represent 3D points otherwise. The two red cameras represent the pose of the two views. The red line indicates the displacement baseline.
• Figure 2: Noise test results of initial relative pose estimation methods.
• Figure 3: Noise test results of optimization methods, which are all initialized by LiRP.
• Figure 4: Robustness test for different residuals. All variants of MS-TLS utilize LiRP for relative rotation. The dashed line represents the benchmark result (RANSAC+5pt) provided by OpenGV.
• Figure 5: Accuracy of the GNC-RANSAC fusion scheme as compared with state-of-art methods. LiGTopt is optimized based on our GNC-RANSAC scheme, and OpenGVopt is optimized based on OpenGV's RANSAC+5pt. The first and second columns are for normal scenes and planar scenes, respectively. Dashed and solid lines represent the initial and optimization relative rotation results, respectively.