Linear Relative Pose Estimation Founded on Pose-only Imaging Geometry
(2401.13357)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.
Overview
-
A new linear relative pose estimation algorithm, LiRP, has been introduced, leveraging advances in pose-only imaging geometry.
-
LiRP integrates a linear global translation constraint into existing methods, improving robustness and accuracy, especially in handling planar degenerate scenes and high outlier ratios.
-
Unlike traditional methods, LiRP identifies outliers using pose-only imaging geometry, decoupling 3D scene structure from camera pose estimation.
-
Extensive testing, including simulations and empirical studies on the Strecha dataset, validates LiRP's effectiveness in maintaining relative rotation accuracy amid significant levels of noise and outliers.
-
The study suggests future research directions in refining the algorithm to minimize parameter tuning and enhance 3D visual computing applications.
Introduction
In the exploration of efficient and robust two-view relative pose estimation, one of the foremost computational challenges faced pertains to the handling of outliers in image matching. Traditional robust estimation methods like RANSAC have certain limitations, such as the requirement for a majority of minimal point pairs to be inliers. Addressing these challenges, a new linear relative pose estimation algorithm has been developed, utilizing recent advances in pose-only imaging geometry. This innovative approach facilitates the processing of a batch of point pairs simultaneously (with n≥6), thereby enhancing accuracy while handling planar degenerate scenes and substantial outlier ratios effectively.
Algorithm Overview
At the core of this algorithm lies the integration of the linear global translation (LiGT) constraint into iterative reweighted least squares (IRLS) and RANSAC methods, which leads to a significant improvement in the robustness of outlier removal. The algorithm, dubbed LiRP (Linear Relative Pose), differs from classical approaches by using pose-only imaging geometry to identify outliers instead of relying on point correspondences to solve the essential matrix. This decouples the 3D scene from camera pose and allows the algorithm to solve for global translations linearly, thus it can handle specific degenerate motions, which classical methods struggle with.
Experimental Validation
Through extensive simulation and empirical tests, including those on the Strecha dataset, the LiRP algorithm demonstrated its capacity to maintain high accuracy in relative rotation estimation – even with an 80% presence of outliers. It showed a noticeable improvement over established methods like the Stewenius five-point algorithm and others supported by OpenGV, especially in handling planar scenes. Importantly, noise tests confirmed the robustness of LiRP, showcasing its competency in maintaining rotation error lower than one degree amid increasing noise.
Conclusion and Further Work
The proposed LiRP algorithm represents a significant stride forward in relative pose estimation under the challenge of outlier management. Its ability to provide a robust, accurate framework for such scenarios is underpinned by well-structured simulation and real-world results. Notably, this work illuminates a path for further research on refinement strategies that could potentially eliminate the need for parameter tuning inherent in methods like IRLS and RANSAC, further propelling advancements in 3D visual computing applications like SLAM and SfM.
Create an account to read this summary for free: