Research map / Point clouds / Segmentation and detection
Fast Point Transformer
Chunghyun Park, Yoonwoo Jeong, Minsu Cho, Jaesik Park
Check 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 Fast Point Transformer instead of Point Transformer | 3D semantic segmentation · S3DIS Area 5 test | Latency (seconds): 18.07 -> 0.14 | Improved |
| use Fast Point Transformer instead of MinkowskiNet42† | 3D semantic segmentation · S3DIS Area 5 test | mIoU (%): 67.1 -> 68.5 | Improved |
| use Fast Point Transformer instead of MinkowskiNet42 | 3D semantic segmentation · ScanNet validation, voxel size 2cm | mIoU (%): 72.2 -> 72.1 | No clear effect |
| replace MinkowskiNet† backbone with Fast Point Transformer backbone | VoteNet · ScanNet validation | mAP@0.25: 55.3 -> 59.1 | Improved |
| add centroid-aware voxelization encoding (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 59.1 -> 61.3 | Improved |
| add centroid-aware devoxelization encoding (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 61.3 -> 62.7 | Improved |
| add continuous positional encoding δ_abs (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 62.7 -> 65.3 | Improved |
| use cosine attention instead of softmax attention (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 61.0 -> 65.3 | Improved |
| add query-key similarity to cosine attention (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 65.3 -> 62.1 | Hurt |
| increase local window size from 3 to 5 (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 65.3 -> 62.4 | Hurt |
| increase local window size from 3 to 7 (ablation) | Fast Point Transformer · ScanNet validation, voxel size 10cm | mIoU (%): 65.3 -> 61.9 | Hurt |
| decompose continuous positional encodings (ablation) | Fast Point Transformer · ScanNet validation, voxel size 2cm | Peak memory usage (GB): 9.52 -> 3.61 | Improved |
Closest papers by meaning
- Point Transformer (ICCV 2021 · similarity 0.91)
- Voxel Set Transformer: A Set-to-Set Approach to 3D Object Detection From Point Clouds (CVPR 2022 · similarity 0.89)
- Spatial Transformer for 3D Point Clouds (TPAMI 2021 · similarity 0.89)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.89)
- LitePT: Lighter Yet Stronger Point Transformer (CVPR 2026 · similarity 0.88)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.88)
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.
