Train in Germany, Test in The USA: Making 3D Object Detectors Generalize

May 2020

tl;dr: The main obstacle to generalized 3D object detection is the size stats of 3D objects in different locations.

Overall impression

This is another insightful work from the same authors of Pseudo-lidar and Pseudo_lidar++. The paper founds out that the size stats affect the localization performance (IoU=0.7) of 3D object detector. However detection at lower IoU thresholds does not degrade much.

The paper converted recently released lidar datasets (KITTI, nuScenes, Argoverse, Lyft, Waymo) into the format of KITTI, including car/truck taxonomy, truncation/occlusion level, difficulty levels. This is quite engineering heavy work.

Key ideas

Technical details