SuperPoint: Self-Supervised Interest Point Detection and Description

摘要: This paper presents a self-supervised framework for training interest point detectors and descriptors suitable large number of multiple-view geometry problems in computer vision. As opposed to patch-based neural networks, our fully-convolutional model operates on full-sized images jointly computes pixel-level locations associated one forward pass. We introduce Homographic Adaptation, multi-scale, multi-homography approach boosting detection repeatability performing cross-domain adaptation (e.g., synthetic-to-real). Our model, when trained the MS-COCO generic image dataset using is able repeatedly detect much richer set points than initial pre-adapted deep any other traditional corner detector. The final system gives rise state-of-the-art homography estimation results HPatches compared LIFT, SIFT ORB.