Robust Keypoint Detection and Matching on Fisheye Images by Self-Supervised Learning

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Accurate feature point detection and matching are crucial for computer vision applications like panoramic stitching and 3D reconstruction.
• Standard feature point methods fail on fisheye images due to severe distortion, rendering conventional camera models inadequate.

Purpose of the Study:

• To propose a novel self-supervised learning method for robust feature point detection and matching specifically designed for highly distorted fisheye images.
• To address the limitations of existing methods in handling the unique challenges presented by fisheye lens imagery.

Main Methods:

• A Siamese network architecture is employed for self-supervised learning, enabling automatic correspondence discovery between transformed image pairs and reducing annotation dependency.
• A two-stage viewpoint transformation pipeline is utilized for data augmentation to mitigate the scarcity of fisheye image datasets.
• Deformable convolution and contrastive learning loss are incorporated to enhance feature extraction and description in distorted image areas.

Main Results:

• The proposed method demonstrates superior performance in feature point detection and matching on fisheye images when compared to traditional approaches.
• Self-supervised learning effectively learns feature point correspondences, overcoming the need for extensive manual labeling.

Conclusions:

• The developed self-supervised learning framework offers a significant advancement for feature point processing in fisheye imagery.
• This method provides a robust and efficient solution for computer vision tasks involving wide-angle and distorted visual data.