RadarPillars: Efficient Object Detection from 4D Radar Point Clouds
Paper • 2408.05020 • Published
RadarPillars — View-of-Delft (radar-only 3D object detection)
Radar-only 3D object detection on the View-of-Delft (VoD) dataset — an OpenPCDet-based reproduction of RadarPillars (Musiat et al., IROS 2024). This checkpoint reproduces and exceeds the published result using 4D radar point clouds only (no camera, no LiDAR).
- 📄 Paper: https://arxiv.org/abs/2408.05020
- 💻 Code, configs & full ablation: https://github.com/fthbng77/RadarPillar
Results (mAP_3D, R11)
| Method | Car @ 0.50 | Ped @ 0.25 | Cyc @ 0.25 | mAP_3D |
|---|---|---|---|---|
| This checkpoint (best seed, s3) | 41.58 | 44.78 | 71.31 | 52.56 |
| 3-seed mean | 41.02 | 43.15 | 70.12 | 51.43 ± 0.99 |
| RadarPillars (paper) | 41.10 | 38.60 | 72.60 | 50.70 |
+1.86 mAP_3D over the paper (best seed). Checkpoint: seed s3 @ epoch 60 (eval-best).
Usage
git clone https://github.com/fthbng77/RadarPillar
cd RadarPillar
python setup.py develop
# download this checkpoint
huggingface-cli download fthbng77/radarpillars-vod radarpillar_vod_best_map52.56.pth --local-dir weights
# evaluate
python tools/test.py \
--cfg_file tools/cfgs/vod_models/vod_radarpillar_rot.yaml \
--ckpt weights/radarpillar_vod_best_map52.56.pth
Architecture
PillarVFE (radial-velocity decomposition) → PillarAttention (masked self-attention) → PointPillarScatter → BaseBEVBackbone → AnchorHeadSingle (Car / Pedestrian / Cyclist). ~0.27M params. See the GitHub repo for full details.
Citation
@inproceedings{radarpillars2024,
title = {RadarPillars: Efficient Object Detection from 4D Radar Point Clouds},
author = {Musiat, Alexander and Reichardt, Laurenz and Schulze, Michael and Wasenm{\"u}ller, Oliver},
booktitle = {Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems (IROS)},
year = {2024}
}
License: Apache-2.0. Built on OpenPCDet.