Research map / Detection and segmentation / 3D detection and driving
RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye View for 3D Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Systematically analyzes radar-camera robustness under simulated radar and image corruptions and proposes RobuRCDet, with a 3D Gaussian Expansion module for noisy radar points and a confidence-guided cross-attention fusion for camera degradation.
Problem
Radar-camera BEV 3D object detection degrades under camera degradation (rain, fog, snow, night) and radar noise (position, RCS, velocity errors), yet robustness has been little studied and no radar-camera corruption benchmark exists.
Approach
A 3DGE module spreads radar RCS and Doppler values over neighboring voxels with a learned adaptive Gaussian kernel (driven by RCS/velocity priors) so dense target regions are reinforced and isolated false points suppressed; a CMCA module predicts a camera confidence map to adaptively weight image vs. radar features in BEV fusion.
Key results
- On nuScenes val with ResNet-50, RobuRCDet reaches 56.7 NDS and 51.2 mAP, versus RCBEVDet's 56.8 NDS and 45.3 mAP (higher mAP than CRN's 49.0).
- With ResNet-18, RobuRCDet reports 55.0 NDS and 45.5 mAP, improving mAP over RCBEVDet (42.9) and CRN (44.9).
- Under the synthesized corruption condition, 3DGE raises NDS from 28.5 to 33.6 and mAP from 23.9 to 29.4 over the radar-camera baseline (ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 3DGE (ablation) | RobuRCDet radar-camera detector · nuScenes val, normal condition | NDS: 54.3 -> 54.9 (+0.6) | Helped |
| add 3DGE (ablation) | RobuRCDet radar-camera detector · nuScenes val, corruption condition | NDS: 28.5 -> 33.6 (+5.1) | Helped |
| add CMCA (ablation) | RobuRCDet radar-camera detector · nuScenes val, normal condition | NDS: 54.3 -> 55.2 (+0.9) | Helped |
| add CMCA (ablation) | RobuRCDet radar-camera detector · nuScenes val, corruption condition | NDS: 28.5 -> 33.1 (+4.6) | Helped |
| use a uniform rather than adaptive kernel map (ablation) | 3DGE · nuScenes val, clean | mAP: 45.5 -> 44.0 | Hurt |
| include x, y, and z in the adaptive kernel parameter encoder (ablation) | 3DGE · nuScenes val, clean | mAP: 45.5 -> 42.7 | Hurt |
| use adaptive 3DGE with RCS and velocity priors rather than the baseline (ablation) | RobuRCDet · nuScenes val, clean | mAP: 44.2 -> 45.5 | Helped |
| use adaptive 3DGE with RCS and velocity priors rather than uniform 3DGE (ablation) | RobuRCDet · nuScenes val, clean | mAP: 44.0 -> 45.5 | Helped |
| replace CRN's fusion module with CMCA (ablation) | CRN · nuScenes val, rainy | NDS: 56.1 -> 57.5 (+1.4) | Helped |
| replace CRN's fusion module with CMCA (ablation) | CRN · nuScenes val, rainy | mAP: 47.3 -> 48.0 (+0.7) | Helped |
| use RobuRCDet instead of CRN | radar-camera detector · nuScenes val, snow | NDS: 41.6 -> 44.7 | Helped |
| use RobuRCDet instead of CRN | radar-camera detector · nuScenes val, snow | mAP: 30.1 -> 32.8 | Helped |
Surprising findings
- The uniform-kernel 3DGE and the variant using position (xyz) inputs to the kernel encoder both reduce performance relative to the baseline, indicating that position is not useful for deciding false positives.
Limitations
- Corruptions are simulated on nuScenes rather than collected in real adverse conditions
- Real-world rainy/night evaluation uses nuScenes subsets with modest gains over baselines
- Performance on clean NDS is not uniformly improved (e.g., RCBEVDet higher NDS with ResNet-50)
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes val
Metrics: NDS mAP mATE mASE mAOE mAVE mAAE
Baselines: CenterFusion CRAFT RCBEV RCBEVDet CRN BEVDet BEVDepth SOLOFusion StreamPETR RCM-Fusion
Builds on: CRN BEVDepth RCBEVDet
Closest papers by meaning
- RCBEVDet: Radar-camera Fusion in Bird's Eye View for 3D Object Detection (CVPR 2024 · similarity 0.92)
- CRT-Fusion: Camera, Radar, Temporal Fusion Using Motion Information for 3D Object Detection (NeurIPS 2024 · similarity 0.90)
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.89)
- RaCFormer: Towards High-Quality 3D Object Detection via Query-based Radar-Camera Fusion (CVPR 2025 · similarity 0.87)
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.87)
- L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.86)
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.