Gaussian YOLOv3: An Accurate and Fast Object Detector Using Localization Uncertainty for Autonomous Driving

November 2019

tl;dr: A probabilistic object detector with NLL loss to maximize the probability of labels.

Overall impression

This paper builds on top of YOLOv3 and makes each of the x, y, w, h two regression target, one mean and one stdev.

FP from a false localization (phantom tracks) during autonomous driving can lead to fatal accidents and hinder safe and efficient driving.

Adding the uncertainty prediction helps with TP and reduce FP dramatically.

The NLL loss essentially models aleatoric uncertainty.

Key ideas

Technical details