Research map / Point clouds / Point cloud classification
PointConvFormer: Revenge of the Point-Based Convolution
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 PointPWC-Net feature-pyramid PointConv layers with PointConvFormer | PointPWC-Net · FlyingThings3D | EPE3D (m): 0.0588 -> 0.0416 | Helped |
| Replace PointPWC-Net feature-pyramid PointConv layers with PointConvFormer | PointPWC-Net · KITTI | EPE3D (m): 0.0694 -> 0.0479 | Helped |
| Use feature-difference attention instead of no attention (ablation) | PointConvFormer-Lite · ScanNet, 5cm grid | mIoU (%): 71.1 -> 73.3 | Helped |
| Use feature-difference attention instead of Point Transformer attention (ablation) | PointConvFormer-Lite · ScanNet, 5cm grid | mIoU (%): 71.6 -> 73.3 | Helped |
| Use feature-difference attention instead of dot-product attention (ablation) | PointConvFormer-Lite · ScanNet, 5cm grid | mIoU (%): 72.0 -> 73.3 | Helped |
| Increase neighborhood size from 4 to 8 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 64.61 -> 69.54 | Helped |
| Increase neighborhood size from 8 to 16 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 69.54 -> 71.40 | Helped |
| Increase neighborhood size from 16 to 32 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 71.40 -> 71.19 | No clear effect |
| Increase neighborhood size from 16 to 48 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 71.40 -> 69.84 | Hurt |
| Use 8 attention heads instead of 2 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 70.71 -> 71.40 | Helped |
| Use 8 attention heads instead of 4 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 70.58 -> 71.40 | Helped |
| Use 16 attention heads instead of 8 (ablation) | PointConvFormer · ScanNet, 10cm grid | mIoU (%): 71.40 -> 70.97 | Hurt |
Closest papers by meaning
- PointConv: Deep Convolutional Networks on 3D Point Clouds (CVPR 2019 · similarity 0.93)
- PointCNN++: Performant Convolution on Native Points (CVPR 2026 · similarity 0.91)
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation (CVPR 2017 · similarity 0.89)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.88)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.87)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.87)
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.