Research map / Label-efficient and robust learning / Self-supervised and contrastive
Bootstrapping Autonomous Driving Radars with Self-Supervised Learning
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.
Radical pre-trains a radar backbone with self-supervised intra-modal (radar-to-radar) and cross-modal (radar-to-vision) contrastive losses, plus a radar-specific MIMO augmentation, then fine-tunes for radar-only car bounding box detection.
Problem
Annotating automotive radar heatmaps is costly and hard for humans, so supervised radar object detection is limited by scarce labels; the paper asks how to exploit large unlabeled radar data for radar-only 2D BEV car detection.
Approach
Joint embedding distillation from a frozen CLIP vision encoder into a radar backbone, combining radar-to-radar and radar-to-vision contrastive losses, with a new RMM augmentation that masks MIMO virtual antennas and applies random phase noise before range-azimuth integration.
Key results
- Radical fine-tuned improves mAP from 56.5 (Radatron supervised) to 62.3, a 5.8% gain.
- With frozen backbone, Radical reaches 52.6 mAP vs 22.1 for Radatron.
- With 1% labeled data, Radical reaches 39.1 mAP vs 27.9 for Radatron.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add combined intra-modal and cross-modal self-supervised pre-training | Radatron · Radatron test | mAP: 56.5 ± 0.2 -> 62.3 ± 0.6 | Helped |
| Add self-supervised pre-training | Radatron · Radatron test, 10% labeled data | mAP: 42.8 -> 50.2 | Helped |
| Add self-supervised pre-training | Radatron · Radatron test, 1% labeled data | mAP: 27.9 -> 39.1 | Helped |
| Add Radar MIMO Mask augmentation (ablation) | Radical · Radatron test | mAP | Helped |
| Add polar rotation augmentation (ablation) | Radical · Radatron test | mAP | Helped |
| Add center cropping augmentation (ablation) | Radical · Radatron test | mAP | Helped |
| Add horizontal flip augmentation (ablation) | Radical · Radatron test | mAP | Helped |
| Add thresholding augmentation (ablation) | Radical · Radatron test | mAP | Hurt |
| Add Cutout augmentation (ablation) | Radical · Radatron test | mAP | Hurt |
| Add vertical flip augmentation (ablation) | Radical · Radatron test | mAP | Hurt |
| Add all four augmentations: RMM, center cropping, horizontal flip, and rotation (ablation) | Radical · Radatron test | mAP: 61.6 ± 0.7 | Mixed |
| Remove RMM from the four-augmentation combination (ablation) | Radical · Radatron test | mAP: 61.6 ± 0.7 -> 61.3 ± 0.8 | Hurt |
Surprising findings
- Naive vision augmentations such as vertical flip, cutout, and thresholding hurt performance during radar pretraining.
- Directly applying popular SSL frameworks to radar heatmaps yields shortcut representations, as noted from prior work.
Limitations
- Evaluated only on a single radar dataset (Radatron) and only on car detection in 2D BEV.
- Augmentation hyperparameters (p, alpha) were tuned on the test-set evaluation sweeps.
- Fog evaluation is a qualitative controlled experiment only.
Experimental setup
Datasets: Radatron
Benchmarks: Radatron
Metrics: mAP AP50 AP75
Baselines: Radatron Intra-modal CL Cross-modal CL
Builds on: SimCLR MoCo Radatron CLIP
Closest papers by meaning
- Enhancing End-to-End Autonomous Driving with Latent World Model (ICLR 2025 · similarity 0.85)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.84)
- RadSimReal: Bridging the Gap Between Synthetic and Real Data in Radar Object Detection With Simulation (CVPR 2024 · similarity 0.83)
- RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar (NeurIPS 2024 · similarity 0.83)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.83)
- Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations (CVPR 2024 · similarity 0.83)
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.