A Comparison of Feature Detectors for Underwater Sonar Imagery

In this work we compare the performance of seven popular feature detection algorithms on a synthetic sonar image dataset. The dataset consists of a single mine-like object (MLO) superimposed on three different backgrounds: grass, sand ripple, and sand. We explore the performance of Harris, Shi-Tomasi, SIFT, SURF, STAR, FAST, and ORB on each of these backgrounds, and all the backgrounds at once by training an SVM classifier. Performance is evaluated with ROC curves by comparing the number of correctly identified features belonging to objects (True Positives) and the number of incorrectly identified features belonging to background noise (False Positives).