Research map / Point clouds / Point cloud classification
Rethinking Radar 3D Perception: Representation Learning from Raw Spectra
Miao Zhang, Mohamed Abdelsamad, Bin Yang, Sherif Abdulatif, Benedikt Loesch, Marco Altmann, Abhinav Valada
Read the paperCheck your idea against 137,217 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.
Challenges the view that raw radar spectra are too noisy for direct learning, arguing the missing piece is a learned representation rather than the raw data itself.
Problem
Radar 3D perception typically relies on handcrafted reductions, sparse point clouds, or compact Doppler descriptors because raw radar spectra are viewed as too cluttered and noisy for direct representation learning.
Approach
Argues the main limitation is the absence of a learned representation over raw spectra rather than the spectra themselves, and proposes learning representations directly from raw radar spectra for 3D perception. Full method details are not available from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace point cloud with dense RAE/RAE-d spectra under supervised training | PTv3 · K-Radar official split | mAP AP_BEV: 49.0 -> 47.4 (RAE), 45.8 (RAE-d) | Hurt |
| add SPMAE/L-SPMAE self-supervised pretraining on spectra | PTv3 · K-Radar official split | mAP AP_BEV: 48.9 -> 55.9 (RAD) | Improved |
| add spectral leakage corruption to random masking (ablation) | SPMAE · K-Radar official split | mAP AP_BEV: 54.3 -> 55.9 | Improved |
| leakage corruption alone instead of random masking (ablation) | SPMAE · K-Radar official split | mAP AP_BEV: 54.3 -> 53.9 | No clear effect |
| use FOCUS power-based masking instead of global random masking (ablation) | SPMAE · K-Radar official split | mAP AP_BEV: 54.3 -> 52.1 | Hurt |
| stack Doppler as channels (RAE-d) instead of mean Doppler (RAE) (ablation) | L-SPMAE · K-Radar official split | mAP AP_BEV: 53.7 -> 52.2 | Hurt |
| use RAD (x,y,D) Doppler axis instead of RAE (ablation) | L-SPMAE · K-Radar official split | mAP AP_BEV: 53.7 -> 55.9 | Improved |
| use radar-only L-SPMAE instead of LiDAR-guided pretraining (ablation) | PTv3 · K-Radar official split | mAP AP_3D: 46.8 -> 50.1 | Improved |
| train from scratch on L-SPMAE reconstructed spectra (ablation) | PTv3 · K-Radar official split | AP_BEV: 48.9 -> 52.0 | Improved |
| swap backbone to Sparse Encoder (ablation) | L-SPMAE · K-Radar official split | AP_BEV: 47.2 -> 53.1 | Improved |
| pretrain with L-SPMAE with 25% labels (ablation) | PTv3 · K-Radar official split | mAP AP_BEV: 48.9 (100% scratch) vs 50.4 (25% + L-SPMAE) | Improved |
| add SPMAE/L-SPMAE pretraining | PTv3 · RADIal test | mAP: 89.0 -> 90.0 | Improved |
Closest papers by meaning
- AdaRadar: Rate Adaptive Spectral Compression for Radar-based Perception (CVPR 2026 · similarity 0.88)
- RaLD: Generating High-Resolution 3D Radar Point Clouds with Latent Diffusion (AAAI 2026 · similarity 0.88)
- RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar (NeurIPS 2024 · similarity 0.83)
- Bootstrapping Autonomous Driving Radars with Self-Supervised Learning (CVPR 2024 · similarity 0.83)
- NEAR: Neural Electromagnetic Array Response (ICML 2025 · similarity 0.83)
- Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation (NeurIPS 2026 · 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.
