Research map / Detection and segmentation / 3D detection and driving
RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features
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 RadarDistill, a LiDAR-to-radar knowledge distillation framework (CMA densification, AFD activation-based low-level distillation, PFD proposal-based high-level distillation) that boosts radar-only 3D detection on nuScenes.
Problem
Radar-based 3D object detection is limited by sparse, noisy radar BEV features, and LiDAR features are hard to transfer to radar because of the large density gap between the two modalities.
Approach
Densify radar BEV features with dilation-based cross-modality alignment (CMA), then distill LiDAR teacher features into the radar student selectively: low-level features only in activation-defined active/inactive regions (AFD) and high-level features only inside detection-proposal TP/FN/FP regions (PFD), with LiDAR used only during training.
Key results
- RadarDistill reaches 20.5% mAP and 43.7% NDS on the nuScenes test set for radar-only detection, versus 8.6 mAP / 34.7 NDS for the PillarNet-18 baseline.
- On the nuScenes validation set, the full model improves NDS from 27.3 (no components) to 34.7 with CMA, AFD and PFD; removing CMA drops NDS by 4%.
- Integrated into BEVFusion (radar-camera), RadarDistill raises mAP from 38.3 to 39.6 and NDS from 45.3 to 46.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Distill LiDAR features into radar features with RadarDistill | PillarNet-18 radar detector · nuScenes test | mAP: 8.6 -> 20.5 | Helped |
| Add CMA to PillarNet-18 (ablation) | PillarNet-18 radar detector · nuScenes validation | NDS: 27.3 -> 29.3 | Helped |
| Add AFD after CMA (ablation) | PillarNet-18 radar detector · nuScenes validation | NDS: 29.3 -> 33.7 | Helped |
| Add PFD after CMA and AFD (ablation) | PillarNet-18 radar detector · nuScenes validation | NDS: 33.7 -> 34.7 | Helped |
| Remove CMA (ablation) | RadarDistill · nuScenes validation | NDS: 34.7 -> 30.7 | Hurt |
| Use AFD instead of Complete KD (ablation) | PillarNet-18 + CMA · nuScenes validation | mAP: 8.6 -> 10.9 | Helped |
| Use AFD instead of Gaussian KD (ablation) | PillarNet-18 + CMA · nuScenes validation | mAP: 8.8 -> 10.9 | Helped |
| Use AFD instead of FG/BG KD (ablation) | PillarNet-18 + CMA · nuScenes validation | mAP: 9.6 -> 10.9 | Helped |
| Use PFD instead of FG/BG KD (ablation) | PillarNet-18 + CMA + AFD · nuScenes validation | NDS: 33.7 -> 34.7 | Helped |
| Remove scale normalization from PFD (ablation) | RadarDistill PFD · nuScenes validation | mAP: 8.5 -> 11.2 | Hurt |
| Integrate RadarDistill into radar-camera fusion (ablation) | BEVFusion radar-camera model · nuScenes validation | mAP: 38.3 -> 39.6 | Helped |
Surprising findings
- Removing CMA while keeping AFD and PFD causes a 4% NDS drop, indicating densification is critical for LiDAR-to-radar transfer.
Limitations
- Ablations are run on a 1/7 subset of the training set and the validation split, not the full test protocol.
- Fusion integration yields only modest gains, and the authors note more sophisticated fusion may help.
- Performance on hard classes such as pedestrians, motorcycles and bicycles remains low (e.g., 9.2, 15.3, 0.9 AP).
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: mAP NDS mATE mASE mAOE mAVE mAAE per-class AP
Baselines: Radar-PointGNN KPConvPillars PillarNet-18 CenterFusion FCOS3D MonoDIS CenterNet BEVFusion Complete KD Gaussian KD FG/BG KD
Builds on: PillarNet CenterPoint Inheriting Strategy BEVFusion FGD
Closest papers by meaning
- SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection (AAAI 2025 · similarity 0.91)
- CRKD: Enhanced Camera-Radar Object Detection with Cross-modality Knowledge Distillation (CVPR 2024 · similarity 0.90)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.86)
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.86)
- Sunshine to Rainstorm: Cross-Weather Knowledge Distillation for Robust 3D Object Detection (AAAI 2024 · similarity 0.85)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 2024 · similarity 0.85)
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.