Geometric Wide-Angle Camera Calibration: A Review and Comparative Study

Jianzhu Huai, Yuxin Shao, Grzegorz Jozkow, Binliang Wang, Dezhong Chen, Yijia He, Alper YilmazSensors 2024 · OpenReview

TL;DR

This paper provides a comprehensive review and comparative study of geometric wide-angle camera calibration (GCC) tools. It surveys existing camera models, calibration targets, and algorithms, and then empirically evaluates six popular GCC tools using simulated and real-world data, offering practical guidance for practitioners.

What Problem Does This Paper Address?

Accurate geometric measurements from images are crucial for many applications, from remote sensing to robotics. Wide-angle cameras, including fisheye and omnidirectional lenses, are increasingly used for their broad field of view. However, calibrating these cameras to accurately relate incoming light rays to image pixels is challenging due to their significant distortions. While numerous GCC tools exist, there's a lack of clear guidance on which tools and models are best suited for different wide-angle camera scenarios, leading to potential confusion and duplicated effort for researchers and engineers. This paper aims to bridge this gap by providing a structured overview and empirical comparison of available GCC solutions.

Methods

The authors conducted a thorough survey of existing GCC tools, categorizing them based on camera models (global parametric like pinhole, KB, FOV, DS, UCM, and local generic models), calibration targets (checkerboards, AprilGrids, circle grids, etc.), and calibration algorithms. They then performed an empirical evaluation of six prominent target-based GCC tools: BabelCalib, Basalt, Camodocal, Kalibr, the MATLAB calibrator, and the ROS calibrator. This evaluation involved using both simulated data (generated from real data with added noise) and real-world data captured with an industrial camera fitted with six different lenses, resulting in fields of view ranging from 90° to 194°. The study focused on the accuracy and repeatability of these tools across various camera models and lenses.

Datasets

The paper does not introduce new datasets. Instead, it utilizes real-world data captured by an industrial camera (UI-3251LE-M-GL) equipped with six different lenses, providing a range of fields of view (90° to 194°). Simulated data was generated from this real data by obtaining camera poses and parameters from a trusted tool (TartanCalib/Kalibr) and then projecting target landmarks while adding Gaussian noise.

Key Insights and Ideas

  • Model Choice Matters: The choice of camera model is critical. For cameras with a field of view (FOV) below 120°, the pinhole radial tangential model is often sufficient. However, for wider FOVs (above 100°), models like the KB-8 or Mei/EUCM become necessary, with the KB-8 model generally showing good accuracy for large FOVs.
  • Outlier Handling is Crucial: The presence of outliers (e.g., misdetected corners) significantly impacts calibration accuracy. Tools that effectively handle outliers (e.g., using Huber or Cauchy loss) tend to produce more reliable results.
  • Tool Failures Reveal Intricacies: The study highlighted that even well-established GCC tools can fail under certain conditions, particularly with specific camera models, large distortions, or insufficient data. This points to the complexity of initialization and optimization in GCC algorithms.
  • No Single "Best" Tool: While some tools performed better than others in specific scenarios, there isn't a single tool that excels in all situations. The best choice often depends on the specific camera, target, and desired accuracy.
  • Parameter Redundancy: Some camera models, like the Mei model, suffer from parameter redundancy, leading to large variances in estimated parameters even with reasonable reprojection errors.

Key Results

  • Comparative Performance: The evaluation showed that for pinhole radial tangential models, most tools (BabelCalib, Camodocal, Kalibr, MATLAB, ROS) provided similar results, with BabelCalib showing slightly larger errors due to ignoring tangential distortion.
  • Wide-Angle Performance: For KB-8 models, tools like BabelCalib, Basalt, Camodocal, and Kalibr achieved similar results, generally outperforming the pinhole model for wider FOVs. The ROS calibrator showed some issues with outliers and initialization for certain sequences.
  • Mei/EUCM Models: For Mei/EUCM models, Kalibr generally provided the best estimates, with Camodocal and ROS also performing well. BabelCalib and Basalt, using the EUCM model, showed comparable results to Mei models in some cases.
  • Failure Rates: Several tools exhibited failure rates, particularly with challenging sequences (e.g., high distortion lenses like MTV185). Kalibr and Camodocal were generally robust, while Basalt and the ROS calibrator showed more frequent failures.
  • Accuracy: For successful calibrations, focal lengths and principal points were typically estimated within a few pixels of the true or reference values. RMS reprojection errors were generally low, often below 1 pixel for successful runs.

Limitations and Future Work

Limitations: * The study focused on target-based calibration, and did not extensively cover self-calibration methods or deep learning-based approaches, although these were mentioned in the review. * The evaluation of the ROS calibrator was somewhat limited due to its perceived issues with outlier handling and implementation of certain models. * The paper acknowledges that some tools failed for specific sequences or models, highlighting the challenges in achieving consistent calibration across all scenarios.

Future Work: The authors suggest several promising research directions:

  • Interactive Calibration: Developing tools that provide real-time feedback and quality checks to guide users in selecting optimal data and poses for calibration.
  • Static Calibration: Exploring methods for static calibration that reduce the need for manual movements, potentially through robotic arms or active targets.
  • Reconstruction with Calibration: Integrating calibration directly into reconstruction pipelines, especially those based on bundle adjustment or emerging deep learning techniques, to achieve more robust and accurate results in real-world conditions.

Summary updated

BibTeX
@inproceedings{huai2024geometric,
  author = {Jianzhu Huai and Yuxin Shao and Grzegorz Jozkow and Binliang Wang and Dezhong Chen and Yijia He and Alper Yilmaz},
  title = {{Geometric Wide-Angle Camera Calibration: A Review and Comparative Study}},
  booktitle = {Sensors},
  year = {2024},
  url = {https://openreview.net/forum?id=nVK0a21NwZ},
}
Figure