Leveraging Heteroscedastic Aleatoric Uncertainties for Robust Real-Time LiDAR 3D Object Detection

November 2019

tl;dr: Model aleatoric uncertainty in both RPN and FRN.

Overall impression

This work extends previous work towards safe ad.

After modeling aleatoric uncertainty, performance is boosted almost 9%.

As we are modeling aleatoric uncertainty for each bbox, it is hetereoscedastic by nature.

However they did not report how prediction quality changes with predicted uncertainty.

Key ideas

Technical details