Comparative Study of Feature Matching Algorithms

Abstract

In this paper, a comparative study on eight feature matching algorithms: SIFT, ORB, KAZE, AKAZE, Dense SIFT, DAISY, BRISK and FREAK were presented. Feature matching is an important aspect of computer vision used for key point detection and matching across im- ages to perform tasks such as object recognition. Each importantly has different features and performance metrics, making them suitable for cer- tain uses. I have evaluated these algorithms with two main criteria: the computation time, and the number of key points matched between two images. Based on the experimental results, there appear to be substantial performance differences among these algorithms that reveal key features of their relative strengths and weaknesses. This performance analysis will help choose which algorithm to use based on requirements like speed and accuracy. In future work, I will apply this analysis to different data sets and fur- ther discuss the advantages and disadvantages of each feature-matching algorithm in other application scenarios.

Keywords: Feature Detection • SIFT (Scale-Invariant Feature Trans- form) • ORB (Oriented FAST and Rotated BRIEF) • KAZE • AKAZE (Accelerated- KAZE) • Dense SIFT • DAISY (Dense Adaptive Scale- Invariant Descriptor) • BRISK (Binary Robust Invariant Scalable Key- points) • FREAK (Fast Retina Key point).