Model-Aware Pre-Training for Radial Distortion Rectification.

Area of Science:

• Computer Vision
• Image Processing
• Optics

Background:

• Camera lenses introduce optical aberrations, specifically radial distortion, affecting image quality.
• Existing methods underutilize the geometric prior of radial distortion and lack effective prediction targets for rectification.

Purpose of the Study:

• To develop a novel framework, the Radial Distortion Transformer (RDTR), for accurate and efficient radial distortion rectification.
• To leverage the inherent geometric model of radial distortion for improved feature extraction and rectification.

Main Methods:

• Introduced a unified, model-aware pre-training stage to learn distortion patterns using a shared geometric distortion model.
• Developed a deformation estimation stage utilizing a novel backward warping flow prediction target for rectification.
• Implemented a transformer-based architecture for processing distortion features.

Main Results:

• Achieved state-of-the-art performance in radial distortion rectification on a synthetic dataset.
• Demonstrated real-time processing capabilities for distortion rectification.
• Validated the generalization of the RDTR framework on real-world images.

Conclusions:

• The RDTR framework effectively rectifies radial distortion by utilizing a shared geometric model and a novel backward warping flow.
• RDTR offers a significant advancement in real-time image rectification, outperforming existing methods.
• The proposed method shows strong generalization across different image types.