Research map / Detection and segmentation / 3D detection and driving
Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions
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 a 3D-LRF module that fuses LiDAR and 4D radar voxel features in 3D space and a weather-conditional radar-flow gating network (WRGNet) that uses a pre-trained image weather classifier to modulate radar-to-LiDAR information flow for 3D detection in adverse weather.
Problem
Robust 3D object detection from LiDAR and 4D radar under normal and adverse weather, where prior fusion ignores sensor characteristics and weather and compresses 3D information into BEV or pseudo-images.
Approach
Fuse LiDAR and 4D radar directly in 3D voxel space via attention over k-nearest radar voxels around each non-empty LiDAR voxel, and gate the radar-derived fusion features with a weather-conditioned signal from a pre-trained image weather classifier.
Key results
- On K-Radar at IoU=0.5, AP3D of 45.2 for the proposed method vs 37.8 for LiDAR-only RTNH* (about 19% relative gain over second best).
- At IoU=0.3, the method reaches 74.8 AP3D and 84.0 APBEV on total, outperforming InterFusion (53.0 AP3D) and RTNH (37.4 AP3D).
- Under heavy snow, improvements range from 10% to 43% over second-best scores across metrics; RTNH* is slightly better in some sleet and fog metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse LiDAR and 4D radar using 3D-LRF and weather-conditioned radar gating | 3D object detection model · K-Radar, all weather | AP3D at IoU=0.5: 37.8 -> 45.2 | Helped |
| Add 3D-LRF (ablation) | Ours without 3D-LRF · K-Radar, all weather | AP3D at IoU=0.5: 35.9 (-1.9) -> 45.2 (+7.4) | Helped |
| Add WRGNet (ablation) | Ours without WRGNet · K-Radar, all weather | APBEV at IoU=0.3: 84.6 (+8.1) -> 84.0 (+7.5) | No clear effect |
| Replace 3D-LRF with cross-attention fusion (ablation) | Ours · K-Radar, all weather | AP3D at IoU=0.5: 35.1 (+0.2) -> 45.2 (+10.3) | Hurt |
| Replace 3D-LRF with concatenation fusion (ablation) | Ours · K-Radar, all weather | AP3D at IoU=0.5: 34.9 -> 45.2 | Helped |
Surprising findings
- Cross-attention fusion over LiDAR and radar does not beat LiDAR-only AP3D at IoU=0.5, showing naive fusion can hurt precise localization.
- Direct radar-LiDAR fusion without weather gating trades off performance across weather conditions, with the gated version alleviating this.
Limitations
- Model relies heavily on radar under sleet and fog, causing more false positives and slightly lower results than RTNH* in some metrics.
- Evaluated only on the K-Radar dataset, the sole benchmark with 4D radar in adverse weather.
- Performance on the Sedan class only is reported for the main comparison.
Experimental setup
Datasets: K-Radar
Benchmarks: K-Radar
Metrics: AP3D APBEV
Baselines: PointPillars RTNH RTNH* InterFusion concatenation fusion cross-attention fusion
Builds on: RTNH InterFusion PointPillars focal loss
Closest papers by meaning
- L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.93)
- RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye View for 3D Object Detection (ICLR 2025 · similarity 0.89)
- Sunshine to Rainstorm: Cross-Weather Knowledge Distillation for Robust 3D Object Detection (AAAI 2024 · similarity 0.89)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.88)
- TransLoc4D: Transformer-based 4D Radar Place Recognition (CVPR 2024 · similarity 0.87)
- UAWTrack: Universal 3D Single Object Tracking in Adverse Weather (AAAI 2025 · similarity 0.87)
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.