Research map / Point clouds / Point cloud classification
PAConv: Position Adaptive Convolution With Dynamic Kernel Assembling on Point Clouds
Check 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 MLPs with PAConv | PointNet · ModelNet40 | accuracy: 89.2 -> 93.2; 4.0%↑ | Helped |
| Replace DGCNN EdgeConv with PAConv | DGCNN · ModelNet40 | accuracy: 92.9 -> 93.6; 1.0%↑ | Helped |
| Replace DGCNN EdgeConv with PAConv and use voting | DGCNN · ShapeNet Parts | class mIoU: 82.3 -> 84.6; 2.3%↑ | Helped |
| Replace DGCNN EdgeConv with PAConv and use voting | DGCNN · ShapeNet Parts | instance mIoU: 85.2 -> 86.1; 0.9%↑ | Helped |
| Replace PointNet++ encoder MLPs with PAConv and use correlation loss | PointNet++ · S3DIS Area-5 | mIoU: 57.27 -> 66.58; 9.31%↑ | Helped |
| Add correlation loss to PAConv (ablation) | PAConv on PointNet++ · S3DIS Area-5 | mIoU: 65.63 -> 66.58; 0.95 mIoU | Helped |
| Use full three-axis coordinate information as ScoreNet input rather than a single-axis representation (ablation) | PAConv · S3DIS Area-5 | mIoU: 63.12 -> 65.63; 63.31 -> 65.63; 64.77 -> 65.63 | Helped |
| Normalize all Weight Bank scores jointly with Softmax instead of no normalization, Sigmoid, or max(0, Tanh) (ablation) | PAConv · S3DIS Area-5 | mIoU: 64.28 -> 65.63; 64.91 -> 65.63; 61.95 -> 65.63 | Helped |
| Vary the Weight Bank size from 2 to 4, 8, or 16 matrices (ablation) | PAConv · S3DIS Area-5 | mIoU: 65.05; 64.86; 64.39; 65.63 | Mixed |
| Increase ScoreNet depth from one to two or three layers (ablation) | PAConv · S3DIS Area-5 | mIoU: 64.42 -> 65.29 -> 65.63 | Helped |
| Remove test-time voting (ablation) | PAConv with correlation loss on PointNet++ · S3DIS Area-5 | mIoU: 66.33 -> 66.58 | Hurt |
Closest papers by meaning
- PointConv: Deep Convolutional Networks on 3D Point Clouds (CVPR 2019 · similarity 0.90)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.88)
- PointConvFormer: Revenge of the Point-Based Convolution (CVPR 2023 · similarity 0.87)
- KPConvX: Modernizing Kernel Point Convolution with Kernel Attention (CVPR 2024 · similarity 0.86)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.85)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (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.