Keypoint Detection Based on the Unimodality Test of HOGs

We present a new method for keypoint detection. The main drawback of existing methods is their lack of robustness to image distortions. Small variations of the image lead to big differences in keypoint localizations. The present work shows a way of determining singular points in an image using histograms of oriented gradients (HOGs). Although HOGs are commonly used as keypoint descriptors, they have not been used in the detection stage before. We show that the unimodality of HOGs can be used as a measure of significance of the interest points. We show that keypoints detected using HOGs present higher robustness to image distortions, and we compare the results with existing methods, using the repeatability criterion.