KP2D: Neural Outlier Rejection for Self-Supervised Keypoint Learning

March 2020

tl;dr: Improvement of unsuperpoint with better designed loss and one auxiliary outlier detection task.

Overall impression

This paper is inspired by unsuperpoint. However it implemented multiple improvement which boosted the performance quite a bit.

Although the proposed method does not achieve best performance all the time, it is within reasonable margin of the best performing model variant.

The method achieves SOTA repeatability and good performance in other metrics.

The proxy task of identifying outliers during training is inspired by NG-RANSAC.

Key ideas

Technical details