Research map / Point clouds / Point cloud classification
SP2T: Sparse Proxy Attention for Dual-stream Point Transformer
Jiaxu Wan, Hong Zhang, Ziqi He, Yangyan Deng, Qishu Wang, Ding Yuan, Yifan Yang
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 |
|---|---|---|---|
| add sparse proxy attention with TRB and global fusion to PTv3 (ablation) | PTv3 · ScanNet val | mIoU: 77.51 -> 78.71 | Improved |
| enlarge PTv3 channels by 1.1x to match size (ablation) | PTv3 · ScanNet val | mIoU: 77.51 -> 77.64 | No clear effect |
| use pool-based point-proxy interaction (ablation) | PTv3 · ScanNet val | mIoU: 77.51 -> 77.47 | No clear effect |
| add Table-based Relative Bias to attention PPI (ablation) | SP2T · ScanNet val | mIoU: 77.83 -> 78.32 | Improved |
| add global fusion over proxies (ablation) | SP2T · ScanNet val | mIoU: 78.32 -> 78.71 | Improved |
| use spatial-wise proxy sampling instead of FPS (ablation) | SP2T · ScanNet val | mIoU: 78.04 -> 78.71 | Improved |
| use spatial-wise sampling instead of fixed-size grid sampling (ablation) | SP2T · ScanNet val | mIoU: OOM -> 78.71 | Improved |
| include empty proxies in global fusion (ablation) | SP2T · ScanNet val | mIoU: 78.03 -> 78.71 | Improved |
| increase proxy count from 160 to 320 or 512 (ablation) | SP2T · ScanNet val | mIoU: 78.71 -> 78.68 (320), 78.52 (512) | No clear effect |
| increase association count from 8 to 12 or 16 (ablation) | SP2T · ScanNet val | mIoU: 78.71 -> 77.90 (12), 78.10 (16) | Hurt |
| share TRB across layers (ablation) | SP2T · ScanNet val | mIoU: 78.33 -> 78.71 | Improved |
| plug SP2T onto MinkUNet local fusion | MinkUNet · ScanNet val | mIoU: 72.2 -> 75.7 | Improved |
Closest papers by meaning
- ProxyTransformation: Preshaping Point Cloud Manifold With Proxy Attention For 3D Visual Grounding (CVPR 2025 · similarity 0.88)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.88)
- Dual Focus-Attention Transformer for Robust Point Cloud Registration (CVPR 2025 · similarity 0.87)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.87)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.87)
- On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation (NeurIPS 2025 · 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.
