Research map / Detection and segmentation / 3D detection and driving
VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add two extra sparse down-sampling layers (strides 16, 32) to the sparse CNN backbone (ablation) | VoxelNeXt sparse backbone · nuScenes validation (1/4 train) | mAP: +9.5 mAP (D3 46.7 to D5 56.5) | Helped |
| increase sparse conv kernels to 5x5x5 instead of adding downsampling (ablation) | D3 backbone · nuScenes validation | mAP: +4.9 mAP but latency 225 ms vs 56 ms | Helped |
| sparse height compression of 3D features to 2D sparse head (ablation) | VoxelNeXt · nuScenes validation | latency: 92 ms to 66 ms with mAP 56.3 to 56.2 | Helped |
| spatial voxel pruning with ratio 0.5 during downsampling (ablation) | VoxelNeXt · nuScenes validation | FLOPs and mAP: FLOPs 83.8G to 33.6G with mAP 56.5 to 56.2 | Helped |
| aggressive voxel pruning ratio 0.9 (ablation) | VoxelNeXt · nuScenes validation | mAP: mAP 56.5 to 45.1 | Hurt |
| 3x3 submanifold sparse conv prediction head instead of FC head (ablation) | VoxelNeXt sparse head · nuScenes validation | mAP: +0.6 mAP (56.2 to 56.8), latency 30 to 35 ms | Helped |
| remove sparse max-pool and NMS (no suppression) (ablation) | VoxelNeXt · nuScenes validation | mAP: 33.0 vs 56.2 with max-pool | Hurt |
| sparse max-pool replacing NMS for post-processing (ablation) | VoxelNeXt · nuScenes validation | mAP: 56.0 (NMS) vs 56.2 (max-pool) | No clear effect |
| voxel association (tracking via query voxels) in addition to center association (ablation) | VoxelNeXt tracking · nuScenes tracking validation | AMOTA: +1.1 AMOTA (69.1 to 70.2) | Helped |
| predict detection heads at stride 16 or 32 instead of 8 (ablation) | VoxelNeXt sparse head · nuScenes validation (1/4 train) | mAP: 56.2 at stride 8 vs 52.5 (16) and 49.0 (32) | Hurt |
| replace CenterPoint dense head with VoxelNeXt fully sparse head | nuScenes validation · nuScenes validation | mAP: +0.9 mAP (55.6 to 56.5) and +1.0 NDS, with 4.9 less orientation error | Helped |
| VoxelNeXt-2D (2D sparse backbone and head) vs 3D backbone (ablation) | VoxelNeXt · nuScenes validation (1/4 train) | mAP: -2.8 mAP (56.2 to 53.4); gap narrows with full data (58.7 vs 60.0) | Hurt |
Closest papers by meaning
- VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection (CVPR 2018 · similarity 0.93)
- Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection (AAAI 2021 · similarity 0.92)
- HVNet: Hybrid Voxel Network for LiDAR Based 3D Object Detection (CVPR 2020 · similarity 0.92)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.92)
- FSHNet: Fully Sparse Hybrid Network for 3D Object Detection (CVPR 2025 · similarity 0.89)
- SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object Detection (CVPR 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.