Probabilistic Camera Distortion Correction Using Deep Gaussian Processes

Area of Science:

• Computer Vision
• Machine Learning
• Computational Imaging

Background:

• Accurate lens distortion correction is crucial for applications like 3D reconstruction and image stitching.
• Traditional methods struggle with highly irregular or non-stationary distortion fields common in specialized cameras.
• Low-data, device-specific settings limit the use of large calibration datasets.

Purpose of the Study:

• To develop a novel framework for lens distortion correction using Deep Gaussian Processes (DGPs).
• To model non-linear mappings for image undistortion in cameras with complex distortion patterns.
• To leverage per-pixel predictive uncertainty for identifying unreliable corrected regions.

Main Methods:

• Proposed a framework based on Deep Gaussian Processes (DGPs) to model non-linear distortion.
• Utilized composed latent mappings in DGPs to represent spatially varying behavior.
• Evaluated on three real camera datasets (RPI, Theta, Pillcam) with varying distortion complexity.

Main Results:

• DGP models demonstrated lower normalized collinearity errors on complex distortion datasets (Theta, Pillcam) compared to standard GP and MLP baselines.
• Polynomial calibration was effective for regular radial distortion (RPI dataset).
• DGP/DGP2 training times ranged from 2383.50s to 10092.50s, indicating computational cost for probabilistic modeling.

Conclusions:

• DGPs provide a robust approach for lens distortion correction in challenging, low-data scenarios.
• The per-pixel uncertainty maps generated by DGPs can enhance downstream task reliability.
• DGPs offer a powerful alternative to conventional methods for non-stationary and irregular distortion fields.