Research map / Point clouds / Point cloud classification
Point Transformer
Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, Vladlen Koltun
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 vector attention instead of scalar attention (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 64.6 -> 70.4 | Improved |
| use attention instead of pointwise MLP with max pooling (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 63.7 -> 70.4 | Improved |
| use relative position encoding instead of none (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 64.6 -> 70.4 | Improved |
| use relative instead of absolute position encoding (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 66.5 -> 70.4 | Improved |
| add relative position encoding to both attention and feature branches instead of only attention (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 67.0 -> 70.4 | Improved |
| add relative position encoding to both branches instead of only feature branch (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 68.7 -> 70.4 | Improved |
| add softmax normalization (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 66.5 -> 70.4 | Improved |
| use k=16 instead of k=4 neighbors (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 59.6 -> 70.4 | Improved |
| use k=16 instead of k=8 neighbors (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 67.7 -> 70.4 | Improved |
| use k=16 instead of larger k=32 (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 68.3 -> 70.4 | Improved |
| use k=16 instead of k=64 (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 67.7 -> 70.4 | Improved |
| use scalar attention instead of pointwise MLP (ablation) | Point Transformer · S3DIS Area 5 | mIoU: 61.7 -> 64.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- RandLA-Net — RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds (CVPR 2020) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Fast Point Transformer (CVPR 2022 · similarity 0.91)
- Self-Positioning Point-Based Transformer for Point Cloud Understanding (CVPR 2023 · similarity 0.91)
- Attention Discriminant Sampling for Point Clouds (ICCV 2023 · similarity 0.89)
- Relation3D : Enhancing Relation Modeling for Point Cloud Instance Segmentation (CVPR 2025 · similarity 0.87)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.87)
- Point Transformer V3: Simpler Faster Stronger (CVPR 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.
