Research map / Point clouds / Segmentation and detection
Point-Voxel CNN for Efficient 3D Deep Learning
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 |
|---|---|---|---|
| Replace PointNet MLP layers with PVConv | PointNet on ShapeNet Part · ShapeNet Part test | mIoU: 83.7 -> 86.2 | Helped |
| Replace point-based feature aggregation with PVCNN | ShapeNet Part models · ShapeNet Part test | Latency: 135.8 ms -> 50.7 ms | Helped |
| Use a narrower 0.25×C PVCNN | PointNet on ShapeNet Part · ShapeNet Part test | mIoU: 83.7 -> 85.2 | Helped |
| Reduce voxel resolution to 0.5×R (ablation) | PVCNN · ShapeNet Part test | mIoU: 86.2 -> 85.5 | Hurt |
| Reduce voxel resolution to 0.75×R (ablation) | PVCNN · ShapeNet Part test | mIoU: 86.2 -> 85.7 | Hurt |
| Use trilinear interpolation during devoxelization (ablation) | PVCNN · ShapeNet Part test | mIoU: +0.5 | Helped |
| Use more than 1× voxel convolution in each PVConv (ablation) | PVCNN · ShapeNet Part test | mIoU: -0.6 -> -0.1 | Helped |
| Replace PointNet++ MLP layers with PVConv | PointNet++ on S3DIS · S3DIS Area 5 | mIoU: 57.26 -> 58.98 | Helped |
| Use PVCNN++ instead of PointCNN | S3DIS segmentation · S3DIS Area 5 | Latency: 282.3 ms -> 69.5 ms | Helped |
| Use PVCNN (1×C) instead of 3D-UNet | S3DIS segmentation · S3DIS Area 5 | mIoU: 54.93 -> 56.12 | Helped |
| Replace F-PointNet MLP layers in segmentation and box estimation with PVConv | F-PointNet on KITTI · KITTI val | Pedestrian AP, Moderate: 55.85 -> 64.71 | Helped |
| Replace only instance-segmentation MLP layers with PVConv | F-PointNet on KITTI · KITTI val | Latency: 29.1 ms -> 58.9 ms | Mixed |
Closest papers by meaning
- PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection (CVPR 2020 · similarity 0.92)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.90)
- HVPR: Hybrid Voxel-Point Representation for Single-Stage 3D Object Detection (CVPR 2021 · similarity 0.90)
- Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection (AAAI 2021 · similarity 0.90)
- Efficient 3D Recognition with Event-driven Spike Sparse Convolution (AAAI 2025 · similarity 0.87)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.85)
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.