Research map / Architectures and efficiency / Neural architecture search
PVNAS: 3D Neural Architecture Search with Point-Voxel Convolution
Zhijian Liu, Haotian Tang, Shengyu Zhao, Kevin Shao, Song Han
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use sparse SPVConv at fine voxel size 0.05 instead of coarse dense PVConv (ablation) | outdoor model · SemanticKITTI large scenes | mIoU: 39.0 -> 58.8 | Improved |
| use PVConv instead of SPVConv for small objects (ablation) | PVCNN · ShapeNet | latency speedup: 2.4x measured speedup | Improved |
| use PVConv instead of SPVConv on PartNet (ablation) | PVCNN (0.5xC) · PartNet | mIoU: 2.2% higher accuracy, 3.8x faster | Improved |
| use PVConv instead of SPVConv on indoor scenes (ablation) | PVCNN · S3DIS | mIoU: 1.2-2.1% higher mIoU | Improved |
| replace MLPs in PointNet with PVConv | PointNet · ShapeNet | mIoU: 83.7 -> 86.2 | Improved |
| replace MLPs in PointNet with PVConv | PointNet · S3DIS | mIoU: 43.0 -> 56.1 | Improved |
| add sparse point-voxel convolution and NAS to MinkowskiNet | SPVNAS-B · SemanticKITTI | mIoU: 63.1 -> 66.4 | Improved |
| add 3D-NAS to SPVCNN with 15G MAC constraint (ablation) | SPVNAS-A · SemanticKITTI | mIoU: 65.3 -> 63.7 with 118.6G -> 15.0 MACs | Improved |
| add 3D-NAS to SPVCNN (ablation) | SPVNAS · nuScenes | mIoU: 54.7 mIoU with 23.1 -> 10.6 GMACs | Improved |
| add 3D-NAS to PVCNN (PVNAS-B vs PVCNN 1xC) (ablation) | PVNAS-B · PartNet | mIoU: 47.8 -> 48.0 | Improved |
| use PVCNN in F-PointNet instead of PointNet++ | F-PointNet · KITTI car moderate | AP: 72.0 -> 72.1 | No clear effect |
| replace SECOND sparse encoder with SPVCNN | SECOND · KITTI car moderate | AP: 80.9 -> 81.8 | Improved |
Closest papers by meaning
- Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution (ECCV 2020 · similarity 0.92)
- Point-Voxel CNN for Efficient 3D Deep Learning (NeurIPS 2019 · similarity 0.91)
- LidarNAS: Unifying and Searching Neural Architectures for 3D Point Clouds (ECCV 2022 · similarity 0.87)
- PV-RCNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection (IJCV 2022 · similarity 0.87)
- PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection (CVPR 2020 · similarity 0.87)
- Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · 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.
