Research map / Detection and segmentation / 3D detection and driving
L4DR: LiDAR-4DRadar Fusion for Weather-Robust 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.
Proposes L4DR, a two-stage LiDAR and 4D radar fusion detector (foreground-aware denoising, multi-modal pillar encoder, and an inter/intra-modal backbone with gated fusion) that is robust to simulated fog and real adverse weather.
Problem
LiDAR-based 3D object detection degrades in adverse weather, and fusing LiDAR with 4D radar is hard because of large data quality disparities between the two sensors and differing degrees of sensor degradation across weather conditions.
Approach
Early data-level fusion of LiDAR and 4D radar pillars (MME) after foreground-aware radar denoising (FAD), followed by a parallel inter-/intra-modal backbone whose features are adaptively gated (MSGF) using the fused features to suppress redundancy and adapt to varying sensor degradation.
Key results
- On Vod-Fog at fog level 4, L4DR achieves 28.9 3D mAP versus 8.92 for the LiDAR-only baseline, and improves 3D mAP by up to 20.0% over LiDAR-only across fog levels.
- On K-Radar, L4DR reaches 53.5 total AP_3D (IoU 0.5), surpassing 3D-LRF (45.2) by 8.3 points.
- On the VoD benchmark, L4DR outperforms InterFusion in all categories, e.g. 69.1 Car entire-area AP vs 66.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use L4DR instead of 3D-LRF | K-Radar · K-Radar Total | AP_3D: 45.2 -> 53.5 | Helped |
| Use L4DR instead of InterFusion | VoD LiDAR-radar fusion · VoD Entire Area, Cyclist | AP: 78.5 -> 82.8 | Helped |
| Add MME to the LiDAR-only baseline (ablation) | L4DR · VoD-Fog, moderate difficulty, no fog | 3D mAP: 70.3 -> 77.1 | Helped |
| Add MME to the LiDAR-only baseline (ablation) | L4DR · VoD-Fog, moderate difficulty, fog level 4 | 3D mAP: 8.92 -> 23.4 | Helped |
| Add FAD to MME (ablation) | L4DR · VoD-Fog, moderate difficulty, no fog | 3D mAP: 77.1 -> 78.7 | Helped |
| Add FAD to MME (ablation) | L4DR · VoD-Fog, moderate difficulty, fog level 4 | 3D mAP: 23.4 -> 24.7 | Helped |
| Add IM² to MME + FAD (ablation) | L4DR · VoD-Fog, moderate difficulty, no fog | 3D mAP: 78.7 -> 78.1 | Hurt |
| Add IM² to MME + FAD (ablation) | L4DR · VoD-Fog, moderate difficulty, fog level 4 | 3D mAP: 24.7 -> 26.5 | Helped |
| Add MSGF to MME + FAD + IM² (ablation) | L4DR · VoD-Fog, moderate difficulty, all fog levels | 3D mAP: 78.1 -> 79.8; 76.8 -> 78.8; 62.0 -> 64.7; 52.3 -> 53.3; 26.5 -> 28.9 | Helped |
| Replace MSGF with concatenation (ablation) | L4DR · VoD-Fog, moderate difficulty, fog level 4 | 3D mAP: 28.8 -> 17.7 | Hurt |
| Replace MSGF with Cross-Attn. (ablation) | L4DR · VoD-Fog, moderate difficulty, no fog vs. fog level 4 | 3D mAP: 79.8 -> 77.2; 28.8 -> 30.1 | Mixed |
| Set FAD training and testing thresholds to 0.3 and 0.2 (ablation) | L4DR · VoD-Fog, moderate difficulty, fog level 4 | 3D mAP: 26.46 -> 28.87 | Helped |
Surprising findings
- Many methods score higher in several adverse weather conditions (e.g., overcast, fog) than in normal weather on K-Radar, attributed to differences in labeled object distributions.
- Attention-based fusion alternatives did not effectively handle fluctuating features across weather, while simple concatenation was competitive at clear weather.
Limitations
- The IM2 and MSGF modules add computation and reduce inference efficiency to about 10 FPS.
- The adverse-weather evaluation on VoD relies on simulated fog rather than real adverse-weather data.
- IM2 causes performance drops in fog and snow on K-Radar, attributed to feature redundancy.
Experimental setup
Datasets: VoD K-Radar Vod-Fog (simulated fog on VoD)
Benchmarks: VoD K-Radar Vod-Fog
Metrics: 3D AP BEV AP 3D mAP
Baselines: PointPillars RTNH InterFusion 3D-LRF LXL FUTR3D BEVFusion RCFusion
Builds on: PointPillars InterFusion OpenPCDet
Closest papers by meaning
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.93)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.88)
- Equirectangular Point Reconstruction for Domain Adaptive Multimodal 3D Object Detection in Adverse Weather Conditions (AAAI 2025 · similarity 0.88)
- Sunshine to Rainstorm: Cross-Weather Knowledge Distillation for Robust 3D Object Detection (AAAI 2024 · similarity 0.87)
- TransLoc4D: Transformer-based 4D Radar Place Recognition (CVPR 2024 · similarity 0.87)
- RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye View for 3D Object Detection (ICLR 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.