LaserNet KL: Learning an Uncertainty-Aware Object Detector for Autonomous Driving

November 2019

tl;dr: Estimating uncertainty with KL divergence loss assuming noisy label and leads to better performance than NLL loss assuming perfect label.

Overall impression

Fomulation of bbox: corners, thus all predictions have equal weights.

Assuming zero noise in label and using NLL to regress for uncertainty has undesirable properties such as infinite gradient around optimum. This leads to the examples with low error and low uncertainty having larger gradients than examples with high error and high uncertainty, and thus lead to numerical instability and overfitting.

This paper assumes each label itself is a distribution and we learn the model by minimizing the Kullback_leibler (KL) divergence between the predicted distribution and the label distribution.

It is important for an autonomous vehicle to understand the uncertainty in its detections due to limited sensor data so that it can plan accordingly, e.g., slowing down the vehicle to collect ore data.

The KL divergence loss is more stable and improves the performance on less common objects.

The paper also proposed a heuristic way to estimate noise in label. –> But this seems problematic and contrived.

Key ideas

Technical details