Extrinsic Camera Calibration without a Direct View Using Spherical Mirror

Amit K. AgrawalICCV 2013 · DOI

Introduction

Extrinsic calibration—the process of determining the rotation and translation of a camera relative to a 3D coordinate system—traditionally requires a direct line of sight to a known reference object, such as a checkerboard. However, many practical scenarios preclude a direct view, including non-overlapping multi-camera systems, human-computer interaction via eye reflections, and the reconstruction of specular objects.

While previous research utilized planar mirrors to circumvent this limitation, these methods are often unstable, requiring at least three distinct mirror orientations and suffering from "blind spots" known as degenerate configurations. This paper, presented at ICCV 2013, proposes a mathematically robust alternative: using a single reflection in a spherical mirror of known radius.

Methodology

The core innovation lies in shifting the geometry from planar to spherical. A spherical mirror transforms a standard perspective camera into a non-perspective axial camera. This geometric complexity, rather than being a hindrance, encodes sufficient information into a single view to eliminate the need for multiple mirror placements.

Geometry of the spherical mirror reflection model, illustrating the axial camera setup.

The calibration process follows a three-stage pipeline:

  1. Axis Estimation: Utilizing a "coplanarity constraint," the axis connecting the camera center to the sphere center is estimated linearly using eight point correspondences.
  2. Analytical Solution: The author derives a closed-form analytical solution for the distance to the sphere and the remaining pose parameters by solving a 16th-degree polynomial. This provides a complete set of potential solutions, with the correct one selected by minimizing re-projection error.
  3. Non-linear Refinement: The nine parameters (six for pose, three for sphere location) are optimized using the Levenberg-Marquardt algorithm to minimize final image re-projection error.

Key Insights

  • Geometric Robustness: Unlike planar methods, the spherical approach has no degenerate configurations. The placement of the sphere relative to the camera or object does not risk mathematical failure.
  • The Benefit of Curvature: The curvature of the mirror increases the "non-centrality" of the camera system. Larger spherical mirrors generally yield higher accuracy because they strengthen the geometric constraints against imaging noise.
  • Practical Simplicity: Since only one view is required and the sphere's location is estimated during the process, the physical setup is significantly simpler than managing multiple planar mirror orientations.

Experimental Results

The method was validated through extensive simulations and real-world experiments using a Canon Rebel XT camera and a high-precision 3-inch spherical ball. Performance was compared against the state-of-the-art planar mirror method by Takahashi et al.

Comparative performance analysis showing error metrics for rotation, translation, and sphere location refinement.

The spherical method demonstrated marked superiority in both simulation and practice:

  • Accuracy under Noise: In simulations with 1-pixel Gaussian noise, the spherical method's initial translation error was 11.9%, compared to 55.3% for the planar method. After non-linear optimization, the spherical method achieved a translation error of 2.4%, outperforming the planar method's 6.0%.
  • Real-World Performance: On real data, the spherical method achieved a root-mean-squared (RMS) re-projection error of 0.17 pixels, a significant improvement over the 1.28 pixels produced by the planar method.

Limitations and Future Work

While highly effective, the algorithm requires a known sphere radius. Additionally, because the spherical mirror reflects the entire environment, the reference object occupies a smaller portion of the image, necessitating high-resolution sensors for accurate corner detection. Future work may explore calibrating systems with unknown radii or applying this foundation to complex multi-camera rigs where overlapping views are physically impossible to achieve.

Summary updated

BibTeX
@inproceedings{agrawal2013extrinsic,
  author = {Amit K. Agrawal},
  title = {{Extrinsic Camera Calibration without a Direct View Using Spherical Mirror}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2013},
  doi = {10.1109/iccv.2013.294},
}
Figure