Research map / Detection and segmentation / 3D detection and driving
V2X-R: Cooperative LiDAR-4D Radar Fusion with Denoising Diffusion 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.
Introduces V2X-R, a simulated V2X dataset with LiDAR, camera and 4D radar, and a cooperative LiDAR-4D radar fusion pipeline with a Multi-modal Denoising Diffusion module that denoises weather-corrupted LiDAR features conditioned on 4D radar features.
Problem
Cooperative LiDAR-based 3D object detection in V2X settings degrades in adverse weather because LiDAR and camera are weather-sensitive, and no cooperative 4D radar dataset or LiDAR-4D radar cooperative fusion method existed.
Approach
Treat agent-fused LiDAR BEV features corrupted by weather noise as samples to be denoised by a conditional diffusion model, using the weather-robust 4D radar BEV feature as the condition, with a decaying epoch-dependent MDD loss weight.
Key results
- MDD improves AttFuse under fog by up to 5.73% 3D mAP (IoU=0.7) and L4DR under snow by up to 6.70% on V2X-R, with negligible normal-weather change.
- LiDAR-4D radar fusion raises CoBEVT, CoAlign and Where2comm 3D mAP@testing (IoU=0.7) by 5.86%, 3.68% and 8.23% over LiDAR-only.
- On K-Radar real-world adverse weather, MDD yields 5.20% and 5.97% gains for Sedan and Bus classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 4D radar fusion to AttFuse | AttFuse · V2X-R snow | 3D mAP@IoU=0.7: 45.54 -> 63.71 | Helped |
| add 4D radar fusion to AttFuse | AttFuse · V2X-R fog | 3D mAP@IoU=0.7: 61.52 -> 62.89 | Helped |
| add MDD to AttFuse with LiDAR-4D radar fusion | AttFuse (LiDAR+4D radar) · V2X-R fog | 3D mAP@IoU=0.7: +5.73 | Helped |
| add MDD to AttFuse with LiDAR-4D radar fusion | AttFuse (LiDAR+4D radar) · V2X-R snow | 3D mAP@IoU=0.7: +3.15 | Helped |
| add MDD to AttFuse with LiDAR-4D radar fusion | AttFuse (LiDAR+4D radar) · V2X-R normal | 3D mAP@IoU=0.7: +0.21 | No clear effect |
| add MDD to L4DR | L4DR · V2X-R snow | 3D mAP@IoU=0.7: +6.70 | Helped |
| replace Vanilla-Unet denoising with diffusion denoising (ablation) | AttFuse MDD · V2X-R fog | 3D mAP@IoU=0.7: 62.07 -> 68.08 | Helped |
| add 4D radar conditioning to diffusion denoising (ablation) | AttFuse MDD · V2X-R fog | 3D mAP@IoU=0.7: 68.08 -> 68.64 | Helped |
| use 3 diffusion steps instead of 2 (ablation) | AttFuse MDD · V2X-R fog | 3D mAP@IoU=0.7: 65.92 -> 68.64 | Helped |
| use epoch-varying loss weight γ(e, ψ=3) rather than constant weight 3 (ablation) | AttFuse MDD · V2X-R fog | 3D mAP@IoU=0.7: 65.92 -> 68.14 | Helped |
| add MDD to single-agent AttFuse | AttFuse · K-Radar adverse weather | 3D mAP (Sedan): 69.04 -> 74.03 | Helped |
| add MDD to single-agent AttFuse | AttFuse · K-Radar adverse weather | 3D mAP (Bus): 50.95 -> 54.58 | Helped |
Surprising findings
- The newest method SICP underperforms earlier cooperative LiDAR methods such as CoBEVT, CoAlign and AdaFusion, because it emphasizes preserving individual perception.
- Multi-agent communication amplifies weather noise in LiDAR point clouds as it propagates, making denser noise a key challenge.
Limitations
- Exploration of cooperative LiDAR and 4D radar fusion is described as underdeveloped.
- The 4D radar sensor provides only front-view data, so evaluation is restricted to the ego camera FOV.
- MDD adds about 32 ms inference time per frame.
Experimental setup
Datasets: V2X-R K-Radar
Benchmarks: V2X-R K-Radar
Metrics: 3D mAP at IoU 0.3/0.5/0.7
Baselines: V2XViT AttFuse Where2comm SCOPE CoBEVT CoAlign AdaFusion SICP L4DR InterFusion
Builds on: DDPM OpenCDA AttFuse L4DR InterFusion
Closest papers by meaning
- L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.86)
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.84)
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.82)
- CRKD: Enhanced Camera-Radar Object Detection with Cross-modality Knowledge Distillation (CVPR 2024 · similarity 0.82)
- RCTrans: Radar-Camera Transformer via Radar Densifier and Sequential Decoder for 3D Object Detection (AAAI 2025 · similarity 0.82)
- SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection (AAAI 2025 · similarity 0.82)
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.