Research map / Detection and segmentation / 3D detection and driving
SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar 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 SCKD, a semi-supervised cross-modality distillation framework where a Lidar-radar fusion teacher (with adaptive fusion and modality dropout) distills a radar-only SECOND student via feature (LRFD, FRFD) and pseudo-label output distillation (SSOD).
Problem
Radar-only 3D object detection with 4D radar point clouds suffers from high sparsity and noise, so radar-only detectors underperform; existing cross-modality distillation relies on ground-truth labels and uses teachers with a different modality from the student.
Approach
Use a same-modality (Lidar+radar) fusion teacher with an adaptive fusion module and modality dropout, distill features into a radar-only student via two feature distillation losses with separate adapters, and replace ground-truth supervision with confidence-thresholded teacher pseudo-labels so unlabeled data can be used.
Key results
- On VoD, the radar-only student improves mAP by 10.38% over the SECOND baseline in the entire annotated area (52.08 vs 41.70) and 6.21% in the driving corridor.
- The radar-only student outperforms the prior best radar-only method SMURF by 1.11% mAP in the entire area on VoD.
- On ZJUODset, adding full unlabeled data (SCKD+) raises moderate-level mAP by 5.12% over SECOND (31.68 vs 26.56).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Distill the radar-only SECOND student with SCKD instead of training it with ground-truth labels | SECOND · VoD, entire annotated area | mAP: 41.70 -> 52.08; 10.38% | Helped |
| Distill the radar-only SECOND student with SCKD instead of training it with ground-truth labels | SECOND · VoD, driving corridor | mAP: 65.59 -> 71.80; 6.21% | Helped |
| Train SCKD with more unlabeled data | SCKD · ZJUODset, moderate | mAP: 27.64 -> 31.68 | Helped |
| Replace ground-truth supervision with SSOD (ablation) | SECOND student · VoD, entire annotated area | mAP: 41.70 -> 43.70 | Helped |
| Add SSOD to ground-truth supervision (ablation) | SECOND student · VoD, entire annotated area | mAP: 43.70 -> 45.82 | Helped |
| Remove GT supervision from SSOD + one-adapter FRFD (ablation) | SECOND student · VoD, entire annotated area | mAP: 48.63 -> 50.35; 1.72% | Helped |
| Remove GT supervision from SSOD + one-adapter FRFD (ablation) | SECOND student · VoD, driving corridor | mAP: 68.16 -> 68.59; 0.43% | Helped |
| Use two separate adapters instead of one in FRFD (ablation) | SSOD + FRFD student · VoD, entire annotated area | mAP: 50.35 -> 50.65 | Helped |
| Add LRFD (ablation) | SSOD + FRFD student · VoD, entire annotated area | mAP: 50.65 -> 51.40 | Helped |
| Add GT supervision to SSOD + FRFD + LRFD (ablation) | SECOND student · VoD, entire annotated area | mAP: 52.08 -> 50.48 | Hurt |
| Use a Lidar-radar fusion teacher instead of a Lidar-only teacher (ablation) | SCKD student · VoD, entire annotated area | mAP: 48.24 -> 52.08; 3.84% | Helped |
| Add adaptive fusion to the Lidar-radar teacher (ablation) | SCKD student · VoD, entire annotated area | mAP: 48.64 -> 49.35 | Helped |
Surprising findings
- Using only semi-supervised output distillation (no ground truth) as student supervision outperforms using ground truth together with it when feature distillation is present.
Limitations
- Relies on a teacher trained with labeled data; student performance depends on teacher quality.
- Evaluated only on VoD and ZJUODset; ZJUODset pedestrian class excluded due to limited samples.
Experimental setup
Datasets: VoD ZJUODset
Benchmarks: VoD ZJUODset
Metrics: AP11 mAP AP40 mAP FPS
Baselines: SECOND PointPillars MVFAN RadarPillarNet LXL-R SMURF SBS RadarDistill LXL FUTR3D BEVFusion RCFusion
Builds on: SECOND RadarDistill OpenPCDet mmdetection3d
Closest papers by meaning
- CRKD: Enhanced Camera-Radar Object Detection with Cross-modality Knowledge Distillation (CVPR 2024 · similarity 0.91)
- RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features (CVPR 2024 · similarity 0.91)
- A-Teacher: Asymmetric Network for 3D Semi-Supervised Object Detection (CVPR 2024 · similarity 0.86)
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.86)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.86)
- 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.