Research map / 3D vision / 3D generation
RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar
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 RadarOcc, a 3D semantic occupancy prediction method that operates directly on 4D imaging radar tensors (with Doppler-bin descriptors, sidelobe-aware sparsification, spherical encoding and spherical-to-Cartesian deformable aggregation), benchmarked on K-Radar against LiDAR and camera baselines.
Problem
3D occupancy prediction for autonomous driving has relied on LiDAR or camera, which degrade in adverse weather; radar point clouds discard weak returns and 4D radar tensors are large, noisy, and in spherical coordinates.
Approach
Process the raw 4D radar tensor directly: compress the Doppler axis into per-location statistical descriptors, sparsify per range (sidelobe-aware) instead of globally by percentile, encode features in spherical voxels with sparse convolution and deformable attention, and aggregate to Cartesian occupancy queries via learnable deformable cross-attention to avoid interpolation.
Key results
- RadarOcc achieves 30.4 IoU and 22.6 mIoU at 51.2m on the K-Radar well-condition test split, outperforming radar-based baselines (L-CONet with radar point clouds, 4DRT-XYZ L-baseline).
- RadarOcc is comparable to 16-beam LiDAR and beats monocular and stereo SurroundOcc in most metrics, e.g. 39.5%/19.7% relative mIoU/IoU gains over stereo SurroundOcc at 51.2m.
- An optimized lightweight version reaches 7.46 fps (126% speedup) with minor accuracy loss on a single GTX 3090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 4D radar tensors instead of radar point clouds | RadarOcc · K-Radar well-condition test, 51.2m | IoU (%) | Helped |
| add the Doppler bins descriptor instead of average-pooling the Doppler axis (ablation) | RadarOcc · K-Radar well-condition test, 12.8m | IoU (%): 48.1 -> 48.8 | Helped |
| use sidelobe-aware spatial sparsifying instead of percentile-based sparsifying (ablation) | RadarOcc · K-Radar well-condition test, 12.8m | mIoU (%): 24.1 -> 34.3 | Helped |
| use spherical-based feature encoding and spherical-to-Cartesian feature aggregation instead of Cartesian encoding (ablation) | RadarOcc · K-Radar well-condition test, 12.8m | IoU (%): 46.2 -> 48.8 | Helped |
| add range-wise self-attention (ablation) | RadarOcc · K-Radar well-condition test, 12.8m | FG IoU (%): 18.1 -> 20.7 | Helped |
| vary the number of selected spatial elements per range to Nr=250 rather than using other tested values (ablation) | RadarOcc · K-Radar validation | IoU (%) | Mixed |
| vary the number of reserved Doppler bins (ablation) | RadarOcc · K-Radar validation, 51.2m | IoU (%): 30.9; 28.8; 31.9; 31.1; 30.1 | No clear effect |
| simplify feature encoding and aggregation and convert occupancy decoding from FP32 to FP16 | RadarOcc · K-Radar inference | fps: 3.30 -> 7.46 | Helped |
| simplify feature encoding and aggregation and convert occupancy decoding from FP32 to FP16 | RadarOcc · K-Radar well-condition test, 51.2m | IoU (%): 30.4 -> 29.3 | Hurt |
Surprising findings
- The Doppler bins descriptor gives only marginal gains (row (a) vs (b)), attributed to the K-Radar Doppler measurement range wrapping around and causing velocity ambiguity.
- RadarOcc without optimization runs at only about 3.3 fps, and the authors note the gap to real-time (10 fps).
Limitations
- Single-frame input and output with no temporal modeling or occupancy forecasting.
- Only two classes (foreground and background) due to lack of point-wise semantic annotation.
- Performance is limited at long range and for small or far objects due to low radar resolution and SNR.
Experimental setup
Datasets: K-Radar
Benchmarks: K-Radar well-condition test split
Metrics: IoU mIoU BG IoU FG IoU
Baselines: L-baseline (OpenOccupancy, LiDAR 16/32/64-beam) L-CONet (radar point cloud) L-baseline (4DRT-XYZ) SurroundOcc (camera mono/stereo)
Builds on: OpenOccupancy CONet Deformable DETR SurroundOcc
Closest papers by meaning
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.88)
- Cam4DOcc: Benchmark for Camera-Only 4D Occupancy Forecasting in Autonomous Driving Applications (CVPR 2024 · similarity 0.88)
- UnO: Unsupervised Occupancy Fields for Perception and Forecasting (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)
- Accurate Training Data for Occupancy Map Prediction in Automated Driving Using Evidence Theory (CVPR 2024 · similarity 0.86)
- OctOcc: High-Resolution 3D Occupancy Prediction with Octree (AAAI 2024 · 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.