Comprehensive empirical evaluation of feature extractors in computer vision

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Feature detection and matching are essential for numerous computer vision applications.
• Evaluating traditional feature detectors and descriptors is critical for optimizing performance.
• Existing methods vary in computational efficiency and robustness to image transformations.

Purpose of the Study:

• To comprehensively evaluate traditional feature detection and descriptor algorithms.
• To assess the impact of architectural design and complexity on computational efficiency and robustness.
• To identify the most efficient and robust feature detection and matching methods for diverse image conditions.

Main Methods:

• Analysis of feature extractors: SIFT, SURF, BRIEF, ORB, BRISK, KAZE, AKAZE, FREAK, DAISY, FAST, and STAR.
• Assessment of computational efficiency and robustness under various transformations (rotation, scaling, blurring, etc.).
• Utilized the Image Matching Challenge Photo Tourism 2020 dataset (>1.5 million images) and >2 million comparisons.

Main Results:

• The FAST (Features from Accelerated Segment Test) detector paired with the ORB (Oriented FAST and Rotated BRIEF) descriptor and Brute-Force (BF) matcher demonstrated the highest computational efficiency.
• ORB descriptors showed strong performance on images with affine transformations and brightness changes.
• AKAZE (Accelerated KAZE) exhibited superior robustness against blurring, fisheye distortion, rotation, and perspective distortions.

Conclusions:

• The FAST-ORB-BF combination offers the fastest feature extraction and matching.
• Algorithm choice significantly impacts performance based on specific image distortions and transformations.
• Understanding these trade-offs is key for selecting optimal feature detection and matching strategies in computer vision.