Research map / Detection and segmentation / 3D detection and driving
SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection
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 SAFDNet, a fully sparse LiDAR 3D detector that uses adaptive feature diffusion (AFD) to fill object centers with features while keeping sparsity, matching hybrid detectors on Waymo/nuScenes and beating them on long-range Argoverse2.
Problem
Fully sparse LiDAR 3D object detectors suffer from a center feature missing problem for large objects, and prior sparse methods either use complex multi-stage pipelines or underperform dense hybrid detectors, especially at long range.
Approach
Adaptive feature diffusion selectively expands sparse BEV voxel features within predicted object boxes (range scaled by average object size, via voxel classification), so object centers get features without dense convolutions, and a center-based sparse head is adapted to sparse features.
Key results
- On Argoverse2 (200m range), SAFDNet reaches 39.7 mAP, +2.6 over HEDNet (37.1) at 2.1x its speed and +2.1 over FSDv2 (37.6) at 1.3x speed.
- On Waymo Open test, SAFDNet achieves 79.8 mAPH L1 and 74.6 L2... with 1.3% L2 mAPH gain over FSDv2 on the test set.
- On nuScenes test, SAFDNet achieves 72.3 NDS and 68.3 mAP, best among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace center-voxel heatmap targets with nearest non-empty voxel targets (ablation) | Sparse model derived from HEDNet · Waymo Open | mAPH: 55.2 -> 71.0 | Helped |
| Normalize heatmap targets by each object's maximum value (ablation) | Sparse model with nearest-voxel heatmap · Waymo Open | mAPH: 71.0 -> 71.4 | Helped |
| Add 2D-EDB (ablation) | Normalized-heatmap sparse model · Waymo Open | mAPH: 71.4 -> 71.5 | No clear effect |
| Add adaptive feature diffusion (ablation) | Sparse model with 2D-EDB · Waymo Open | mAPH: 71.5 -> 73.3 | Helped |
| Add feature diffusion (ablation) | SAFDNet sparse detector · Waymo Open | mAPH: 71.5 -> 73.3 | Helped |
| Replace AFD with parameter-based uniform feature diffusion (ablation) | SAFDNet · Waymo Open | mAPH: 73.3 -> 73.0 | No clear effect |
| Replace AFD with parameter-free uniform feature diffusion (ablation) | SAFDNet · Argoverse2 | mAP: 39.7 -> 37.7 | No clear effect |
| Set AFD diffusion-range α to 0.5 instead of 1.0 (ablation) | SAFDNet · Waymo Open | mAPH: 73.3 -> 72.7 | Hurt |
| Set AFD diffusion-range α to 2.0 instead of 1.0 (ablation) | SAFDNet · Waymo Open | mAPH: 73.3 -> 73.0 | Hurt |
| Use SAFDNet instead of FSDv2 | LiDAR 3D object detection · Argoverse2 validation | mAP: +2.1% mAP | Helped |
| Use SAFDNet instead of HEDNet | LiDAR 3D object detection · Argoverse2 validation | mAP: +2.6% mAP | Helped |
| Use SAFDNet instead of FSDv2 | LiDAR 3D object detection · Waymo Open test L2 | mAPH: 1.3% L2 mAPH gains | Helped |
Surprising findings
- Simply replacing dense convolutions with sparse ones in HEDNet drops Waymo mAPH from 73.2 to 55.2 with a center-based Gaussian heatmap, and the nearest-voxel sparse model still trails HEDNet.
- Unrestricted feature diffusion into all unoccupied areas (uniform diffusion) can hurt accuracy; larger alpha (2.0) degraded Waymo mAPH relative to alpha 1.0.
Limitations
- Adaptive diffusion can generate unnecessary feature regions outside objects, reducing sparsity.
- Direction prediction struggles for partially large objects such as box trucks, per qualitative analysis.
Experimental setup
Datasets: Waymo Open nuScenes Argoverse2
Benchmarks: Waymo Open nuScenes Argoverse2
Metrics: mAP mAPH NDS FPS FLOPs training memory
Baselines: HEDNet FSDv1 FSDv2 VoxelNeXt CenterPoint SWFormer DSVT-Voxel LargeKernel3D LinK
Builds on: CenterPoint HEDNet FSDv2 VoxelNeXt OpenPCDet
Closest papers by meaning
- Fully Sparse 3D Object Detection (NeurIPS 2022 · similarity 0.92)
- RSN: Range Sparse Net for Efficient, Accurate LiDAR 3D Object Detection (CVPR 2021 · similarity 0.90)
- LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving (CVPR 2019 · similarity 0.89)
- VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking (CVPR 2023 · similarity 0.89)
- FSHNet: Fully Sparse Hybrid Network for 3D Object Detection (CVPR 2025 · similarity 0.89)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.89)
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.