Reproduce “Benchmarking Robustness of 3D Object Detection to Common Corruptions in Autonomous Driving” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yinpeng Dong, Caixin Kang, Jinlai Zhang, Zijian Zhu, Yikai Wang, Xiao Yang, Hang Su, Xingxing Wei and 1 more

Researchers from Tsinghua University and affiliated institutions established the first comprehensive corruption robustness benchmarks (KITTI-C, nuScenes-C, Waymo-C) for 3D object detection in autonomous driving, introducing 27 realistic corruption types including novel motion-level distortions. Their large-scale evaluation of 24 models revealed that motion-level corruptions are the most detrimental, camera-only models are highly vulnerable, and traditional data augmentation provides limited robustness gains.

Key insights

  • Motion-level corruptions, such as motion compensation and moving object distortions, were identified as the most detrimental across all detector types, highlighting a previously unexplored vulnerability.
  • Camera-only models exhibited extreme vulnerability to common corruptions, with over 40% performance drop, indicating the critical role of LiDAR for reliable 3D perception in adverse conditions.
  • Standard data augmentation strategies provided minimal to no robustness gains, suggesting that enhancing 3D object detection robustness requires more fundamental research into architectures or learning paradigms.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.