Research map / Detection and segmentation / 3D detection and driving
Understanding the Robustness of 3D Object Detection With Bird's-Eye-View Representations in Autonomous Driving
Zijian Zhu, Yichi Zhang, Hai Chen, Yinpeng Dong, Shu Zhao, Wenbo Ding, Jiachen Zhong, Shibao Zheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BEV representation (camera-only BEV models vs non-BEV) | camera-only 3D detectors · ImageNet-C style common corruptions on nuScenes val | mAP/NDS | Improved |
| Adding camera-LiDAR fusion | 3D detectors · common corruptions and weather/lighting | mAP/NDS | Improved |
| Use BEV representation (BEV models vs non-BEV) | camera-only 3D detectors · l_inf FGSM and PGD10 image perturbations | mAP/NDS | Hurt |
| Use Transformer blocks | camera-only detectors · common corruptions | mAP/NDS | Improved |
| Adding camera-LiDAR fusion (TransFusion/BEVFusion vs camera-only) | 3D detectors · l_inf FGSM attack on images | NDS drop at eps=8: BEVFusion drops 10.0 vs 30.6 for BEVFormer | Improved |
| Use BEV representation with temporal history (BEVFormer vs DETR3D) | camera-only detectors · category-specific and temporally universal patch attacks | NDS: BEVFormer outperforms DETR3D by at least 2.2 mAP and 5.8 NDS in temporally universal patch attack | Improved |
| Use BEV representation (BEVDet/BEVDepth vs FCOS3D) under 3D consistent patches | camera-only detectors · instance-specific adversarial patches | mAP/NDS | Hurt |
| Increase input resolution from 704x256 to 1408x512 (ablation) | BEVDet and BEVDepth · l_inf PGD10 attack on nuScenes | NDS: BEVDepth eps=1 NDS 7.0 -> 11.0 (HighRes) | Improved |
| Add temporal information (BEVDet -> BEVDet4D) (ablation) | BEVDet · category-specific patch 5% and temporally universal patch 5% | NDS: 15.4 -> 21.3 (category); 12.7 -> 17.5 (temporal) | Improved |
| Add temporal history (BEVFormer vs BEVFormer-Static) (ablation) | BEVFormer · category-specific patch 5% and temporally universal patch 5% | NDS: 32.9 -> 35.9 (category); 30.3 -> 35.0 (temporal) | Improved |
| Partial cameras: mask 3 non-adjacent cameras (lambda-mode vs Y-mode) | camera-only detectors · nuScenes val partial cameras | NDS | Mixed |
| Adding a fusion model component: BEVFusion vs TransFusion under point cloud perturbation | fusion detectors · l_inf image and point cloud perturbations | robustness | Mixed |
Closest papers by meaning
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (AAAI 2025 · similarity 0.91)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.90)
- BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios (AAAI 2024 · similarity 0.89)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.88)
- RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye View for 3D Object Detection (ICLR 2025 · similarity 0.88)
- BEVDiffuser: Plug-and-Play Diffusion Model for BEV Denoising with Ground-Truth Guidance (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
