Learning to Find Good Correspondences

March 2020

tl;dr: Given the coordinates of matching candidate, assign weight to the candidates.

Overall impression

The paper is inspired by PointNet, and recycles the permutation-invariant property of PointNet to assign weight to correspondence candidates.

The paper uses a hybrid approach of combining classification and estimation of Essential Matrix together. However, even only with classification the performance is not as bad.

Another interesting thing is that the paper uses only the keypoint position of (x1, y1, x2, y2) x Batch_N as input to the PointNet-like neural network and completely discard the descriptors.

This paper inspired NG-RANSAC and KP2D.

Key ideas

Technical details