Reproduce “Deep-BrownConrady: Prediction of Camera Calibration and Distortion Parameters Using Deep Learning and Synthetic Data” on OpenResearch

About this paper

Faiz Muhammad Chaudhry, Jarno Ralli, Jerome Leudet, Fahad Sohrab, Farhad Pakdaman, Pierre Corbani, Moncef Gabbouj

Researchers from AILiveSim Ltd. and Tampere University developed Deep-BrownConrady, a deep learning model that predicts Brown-Conrady camera calibration and distortion parameters from a single image. The model, trained primarily on synthetic data, achieves accuracy comparable to traditional methods requiring five or more calibration images, demonstrating robustness across various image transformations and generalizing to the real-world KITTI dataset.

Key insights

• High-fidelity synthetic data, generated with adaptively determined distortion ranges, can effectively train deep learning models for camera calibration that generalize well to real-world conditions.
• Explicitly feeding normalized image size and aspect ratio information to a deep learning model significantly improves its ability to predict intrinsic camera parameters across varied input resolutions.
• The Brown-Conrady model's distortion effects, particularly for radial components, show a strong relationship with the camera's Field of View, becoming more pronounced with larger FOVs.

Summary by alphaXiv.