Learning to Identify Stable Keypoints

Abstract

Through our experiments with the popular SIFT-DoG keypoint detector, we find that its stability in extracting keypoints from rotated images is good, but sometimes not as good as we expect. This paper presents our endeavor to improve the stability of the DoG keypoint detector by learning from tens of millions of training samples. The learning problem is formulated in a filtering setting, where the training samples are drawn from an oracle instead of using a fixed training set. We show that, by increasing the stability of keypoint detector, we may obtain discriminative local features for matching. The matching accuracy can be improved by 10% using the learned decision function as a watchdog to block unstable keypoints, with acceptable overheads in computation.