Research map / Point clouds / Point cloud classification
Point Transformer V2: Grouped Vector Attention and Partition-based Pooling
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 grouped vector attention (ablation) | PTv1 baseline · ScanNet v2 val | mIoU: 72.3 -> 73.8 | Helped |
| add position encoding multiplier (ablation) | PTv2 · ScanNet v2 val | mIoU: 73.8 -> 74.4 | Helped |
| add grid (partition-based) pooling (ablation) | PTv2 · ScanNet v2 val | mIoU: 74.4 -> 74.9 | Helped |
| add partition map unpooling (ablation) | PTv2 · ScanNet v2 val | mIoU: 74.9 -> 75.4 | Helped |
| GVA instead of multi-head self-attention (ablation) | neighborhood attention · ScanNet v2 val | mIoU: 73.9 -> 75.0 | Helped |
| neighborhood attention instead of shifted-grid attention (ablation) | PTv2 with GVA · ScanNet v2 val | mIoU: 72.5 -> 75.0 | Helped |
| grouped linear instead of linear in weight encoding (ablation) | GVA weight encoding · ScanNet v2 val | mIoU: 74.7 -> 75.0 | Helped |
| grid pooling instead of FPS-kNN pooling at ratio 1/6 (ablation) | PTv2 · ScanNet v2 val | mIoU: 72.9 -> 75.4 | Helped |
| vary grid size multipliers (ablation) | grid pooling · ScanNet v2 val | mIoU: 74.7 to 75.4 | No clear effect |
| deeper decoder blocks (depth 1 -> 2) (ablation) | PTv2 decoder · ScanNet v2 val | mIoU | No clear effect |
| add PE multiplier (ablation) | PTv1 · ScanNet v2 val | mIoU | No clear effect |
Closest papers by meaning
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.90)
- Stratified Transformer for 3D Point Cloud Segmentation (CVPR 2022 · similarity 0.89)
- ConDaFormer: Disassembled Transformer with Local Structure Enhancement for 3D Point Cloud Understanding (NeurIPS 2023 · similarity 0.89)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.87)
- Frozen CLIP Transformer Is an Efficient Point Cloud Encoder (AAAI 2024 · similarity 0.86)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · 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.