Learning-Deep-Learning

LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving

November 2019

tl;dr: A real-time probabilistic lidar detector based on RV (range view) that Models aleatoric uncertainty.

Overall impression

The introduction of probabilistic object detection boosts the performance by nearly 4%. This work is succeeded by LaserNet KL by modeling the noise/uncertainty in label as well.

The paper intentionally only predicts aleatoric uncertainty as there is no efficient way to compute epistemic uncertainty.

The RV view of point cloud is dense and is the native representation of lidar. Projecting to 3D space or BEV leads to sparse representation and more computation.

I feel the authors started with a RV-based lidar detector, but only added probabilistic object detection as a novelty to boost performance.

Key ideas

Technical details

Notes