UnsuperPoint: End-to-end Unsupervised Interest Point Detector and Descriptor

March 2020

tl;dr: Train an interesting point detector without pseudo-gt.

Overall impression

Most previous DL-based interest point detection works focuses on the descriptor part, but still rely on traditional keypoint detection (such as FAST). The main challenge is that the keypoint in natural image is hard to define.

This paper is inspired by SuperPoint. However superpoint requires pseudo-gt generated from synthetic images followed by homographic adapttion (extensive, ~100 TTA to increase recall). Interesting/salient points emerges naturally after training.

This paper has tons of losses! Balancing them is quite a task.

Key ideas

Technical details