NG-RANSAC: Neural-Guided RANSAC: Learning Where to Sample Model Hypotheses

March 2020

tl;dr: Learn the weight of samples (correspondence) for Ransac.

Overall impression

Ransac is an algorithm for robust (insensitive to observation outlier) estimation of parameters. Generally the sampling weights of all observation is uniform. NG-Ransac upweighs correct/in-lier correspondence and downweighs incorrect/outlier correspondence.

The method of telling whether a correspondence (point-pair) is an inlier or not was inspired by learning good correspondence (CVPR 2018) and then both inspired KP2D (ICLR 2020) as a supervision signal. The basic idea is to feed into a PointNet-like structure point-pair candidate and predict the probability of it being an in-lier. –> This can directly benefit the radar pin-camera bbox data association problem.

The paper has a lot of details and is perhaps worth another thorough read.

Key ideas

Technical details