Revisiting Radial Distortion Absolute Pose
Introduction
Estimating a camera's absolute pose—its position and orientation in space—is a fundamental task in computer vision. Standard approaches often assume a pinhole camera model, but real-world lenses, particularly wide-angle ones found on smartphones and action cameras, exhibit significant radial distortion.
Historically, researchers have used "undistortion" models to map distorted pixels back to a pinhole grid. While mathematically convenient because they decouple pose from distortion, these models are less ideal for 3D reconstruction tasks where minimizing reprojection error in the original image space is preferred. This paper bridges this gap by presenting the first minimal solvers that directly handle distortion models, providing a more direct and mathematically consistent way to handle wide-angle lenses in Structure-from-Motion (SfM) and Augmented Reality (AR) pipelines.
Methodology
The proposed framework handles rational distortion models of arbitrary degrees through a clever two-stage pipeline:
- 1D Radial Alignment: The solver first utilizes a "1D radial camera model" to estimate rotation and translation in the XY plane. This stage is effective because the radial direction of points from the image center is invariant to focal length and translation along the principal axis ($t_3$).
- The Upgrade Step: With rotation known, the system solves for the remaining degree of freedom ($t_3$), focal length ($f$), and distortion parameters ($\mu, \lambda$) simultaneously.
The primary challenge in this approach is that the radius of a point in a distortion model depends on its depth ($t_3$), creating a highly non-linear coupling. The authors overcome this using a "hidden variable trick" to isolate $t_3$, transforming the system into either a univariate polynomial or a Polynomial Eigenvalue Problem (PEP), both of which are computationally efficient and stable.
Experimental Results
The solvers were rigorously tested on synthetic data mimicking GoPro characteristics, a real-world checkerboard dataset, and the ETH3D benchmark.
The univariate solvers are exceptionally fast, with execution times ranging from 0.2 to 64 microseconds. Even the more complex PEP-based solvers remain under 111 microseconds, making them ideal for real-time RANSAC loops used to filter out noisy data.
On the ETH3D "Delivery Area" sequence, the $D(3,0)$ model achieved a median position error of 0.12m, matching or exceeding the performance of state-of-the-art undistortion-based methods. The solvers also demonstrated remarkable stability for high-degree models, such as the 6-parameter OpenCV-style model, which was previously considered too complex for a minimal solver.
Discussion and Limitations
A key insight of this work is that the coupling between depth and distortion, previously seen as a barrier, can be managed using modern algebraic geometry tools. This allows for a universal framework applicable to polynomial, division, and complex rational models.
However, the approach faces limitations in fronto-parallel degeneracy. When a camera faces a flat plane directly, it becomes mathematically difficult to distinguish between changes in focal length and forward translation ($t_3$). Additionally, for very complex models with many parameters, there is a risk of overfitting to noise, which can be mitigated by adding a small dampening factor during final refinement.
Conclusion
By revisiting the radial distortion absolute pose problem through the lens of direct distortion models, this paper provides a robust and efficient set of tools for the computer vision community. These minimal solvers offer a more accurate and theoretically sound alternative to traditional undistortion methods for wide-angle camera calibration and localization.
Summary updated