CrashD Dataset: Out-of-Domain Test Benchmark for LiDAR 3D Object Detection

CVPR 2022

Paper: 3D-VField: Adversarial Augmentation of Point Clouds for Domain Generalization in 3D Object Detection

Alexander Lehner*,1,2, Stefano Gasperini*,1,2, Alvaro Marcos-Ramiro2, Michael Schmidt2, Mohammad-Ali Nikouei Mahani2, Nassir Navab1,3, Benjamin Busam1, Federico Tombari1,4
* equal contribution, 1 Technical University of Munich, 2 BMW Group, 3 Johns Hopkins University, 4 Google
CVPR Paper arXiv Paper Video Paper Website

CrashD Download

Part 1 Data Part 2 Data Format Description

CrashD Description

CrashD is a synthetic LiDAR dataset to quantify the generalizability of a 3D object detector on out-of-domain samples. As it includes various types of cars, such as normal, old, sports and damaged, it comprises a variety of plausible vehicle shapes, thereby serving as a valuable out-of-domain test. Specifically, the crashes are individually generated with a realistic simulator and distinguished depending on the intensity, namely light, moderate, hard, as well as the kind of damage: clean (i.e., undamaged), linear (i.e., frontal or rear), and t-bone (i.e., lateral). The randomly and automatically generated 15340 scenes were captured by a 64-beam LiDAR configured to mimic the one in KITTI. Each scene presents between 1 and 5 vehicles, with visible damages, before being repaired and placed at the same locations to collect the clean set, resulting in a total of 46936 cars. We are releasing this data publicly, as an out-of-domain evaluation benchmark for models trained on KITTI, Waymo, or similar datasets. Further details can be found in the Supplementary Material of our CVPR 2022 paper.

Comparison of linear damage intensities for normal and rare cars of CrashD. For each type of car, the accidents were created by the same hitting vehicle, coming from the same angle. It can be seen that the hard crash compromised the structure of the weaker rare car, while the normal car absorbed the impact differently, leaving the cabin unchanged.

BibTeX


@inproceedings{lehner2022_3d-vfield,
  title={{3D-VField}: Adversarial Augmentation of Point Clouds for Domain Generalization in {3D} Object Detection},
  author={Lehner, Alexander and Gasperini, Stefano and Marcos-Ramiro, Alvaro and Schmidt, Michael and Mahani, Mohammad-Ali Nikouei and Navab, Nassir and Busam, Benjamin and Tombari, Federico},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={17295--17304},
  year={2022}
}