Research map / Detection and segmentation / Open-set and few-shot detection
RadSimReal: Bridging the Gap Between Synthetic and Real Data in Radar Object Detection With Simulation
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 RadSimReal, a physical radar simulator that replaces hardware-specific signal processing with convolution of reflection points and a measured point spread function, and shows detectors trained on its synthetic radar images perform comparably to real-data training on RADDet, CARRADA and CRUW.
Problem
Annotated real radar images for training object detection networks are costly to collect and annotate, and existing physical radar simulators need undisclosed radar hardware and signal-processing details and are computationally slow.
Approach
Generate radar tensors by 3D convolution of CARLA-derived reflection points with a truncated radar point spread function (PSF) measured from a single narrow target, plus noise, avoiding hardware design knowledge and reducing computation by about 1000x.
Key results
- U-Net trained on RadSimReal reaches 85.63 AP@0.1 on RADDet test vs 84.76 when trained on RADDet real data.
- Models trained on RadSimReal outperform those trained on CARRADA real data when tested on CARRADA (e.g., U-Net 70.77 vs 50.99 AP@0.1).
- FID between RADDet train set and RadSimReal 10k images is 6.54, vs 6.76 between RADDet train and test sets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train U-Net on RadSimReal instead of RADDet | U-Net · RADDet test | AP@0.5: 55.53 -> 57.64 | Helped |
| Train Probabilistic on RadSimReal instead of RADDet | Probabilistic · RADDet test | AP@0.5: 40.68 -> 52.36 | Helped |
| Train RADDet model on RadSimReal instead of RADDet | RADDet model · RADDet test | AP@0.5: 47.95 -> 46.63 | Hurt |
| Train U-Net on RadSimReal instead of CARRADA | U-Net · CARRADA test | AP@0.5: 44.65 -> 43.96 | Hurt |
| Train Probabilistic on RadSimReal instead of CARRADA | Probabilistic · CARRADA test | AP@0.5: 39.51 -> 41.07 | Helped |
| Train RADDet model on RadSimReal instead of CARRADA | RADDet model · CARRADA test | AP@0.5: 17.79 -> 28.76 | Helped |
| Train U-Net on RadSimReal instead of RADDet | U-Net · CARRADA test | AP@0.5: 30.78 -> 43.96 | Helped |
| Train Probabilistic on RadSimReal instead of RADDet | Probabilistic · CARRADA test | AP@0.5: 19.53 -> 41.07 | Helped |
| Train RADDet model on RadSimReal instead of RADDet | RADDet model · CARRADA test | AP@0.5: 20.49 -> 28.76 | Helped |
| Train U-Net on RadSimReal instead of CRUW | U-Net · CRUW test | AP@0.5: 56.54 -> 55.47 | Hurt |
| Train Probabilistic on RadSimReal instead of CRUW | Probabilistic · CRUW test | AP@0.5: 52.95 -> 51.45 | Hurt |
| Train RADDet model on RadSimReal instead of CRUW | RADDet model · CRUW test | AP@0.5: 54.36 -> 52.48 | Hurt |
Surprising findings
- Models trained on RadSimReal outperform models trained on the real RADDet data when tested on CARRADA, indicating better cross-dataset generalization.
- Adding real RADDet data to RadSimReal training gives no significant improvement, suggesting small sim-to-real domain shift.
Limitations
- Requires a measured radar PSF and noise variance, which needs a narrow reference target measurement.
- Evaluated only on a single class ('car') and TI-prototype-based datasets with azimuth-only radar.
- Generative GAN methods are said to close the gap too but need real data collection; comparison to other physical simulators is not experimentally made.
Experimental setup
Datasets: RADDet CARRADA CRUW RadSimReal (synthetic, generated with CARLA)
Benchmarks: RADDet test set CARRADA test set CRUW test set
Metrics: Average Precision at IoU 0.1, 0.3, 0.5 FID
Baselines: RADDet U-Net Probabilistic GAN-based radar generation
Builds on: CARLA RADDet U-Net radar detection
Closest papers by meaning
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.84)
- Bootstrapping Autonomous Driving Radars with Self-Supervised Learning (CVPR 2024 · similarity 0.83)
- RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features (CVPR 2024 · similarity 0.83)
- NeuRAD: Neural Rendering for Autonomous Driving (CVPR 2024 · similarity 0.82)
- RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation (AAAI 2024 · similarity 0.82)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.82)
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.