PaperReading

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

• The formulation of NG-ransac facilitates training with any non-differentiable task loss function, and any non-differentiable model parameter solver, making it broadly applicable.
• The basic idea of the trick is to formulate the problem as a function of a probability distribution and minimize the expected task loss during training. Add a probabilistic wrapper around a deterministic algorithm to make it partially differentiable.

Technical details

• The idea can be also used for horizon line estimation and camera relocalization.

Notes

• Official Github repo based on pytorch. –> Note that deploy this on Ubuntu. It runs into problems on Mac.