SIR: Self-Supervised Image Rectification via Seeing the Same Scene From Multiple Different Lenses.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning excels at image rectification using supervised methods on synthetic data.
• Supervised models struggle with real-world fisheye images due to overfitting and limited distortion model universality.

Purpose of the Study:

• To develop a self-supervised image rectification (SIR) method.
• To improve generalization of image rectification models on real-world distorted images.

Main Methods:

• A novel network architecture with a shared encoder and multiple prediction heads for distortion parameters.
• A differentiable warping module for generating rectified and re-distorted images.
• Leveraging intra- and inter-model consistency for self-supervised learning.

Main Results:

• Achieved comparable or superior performance to supervised and state-of-the-art methods on synthetic and real-world datasets.
• Demonstrated effective image rectification without ground-truth distortion parameters or normal images.
• Showcased improved universality of distortion models while maintaining self-consistency.

Conclusions:

• The proposed self-supervised image rectification method offers a robust alternative to supervised approaches.
• This method enhances the generalization capability of deep learning models for fisheye image correction.
• The approach provides a pathway for developing more universally applicable distortion models.