Research map / Point clouds / Point cloud classification
Modeling Point Clouds With Self-Attention and Gumbel Subset Sampling
Check 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 |
|---|---|---|---|
| Use FPS for the first down-sampling and GSS for later down-sampling instead of FPS at every stage | PAT classification · ModelNet40 test | accuracy: 91.4 -> 91.7 | Helped |
| Use FPS for the first down-sampling and GSS for later down-sampling instead of FPS at every stage | PAT classification · DVS128 Gesture 10-class | system accuracy: 96.5 -> 97.4 | Helped |
| Use FPS for the first down-sampling and GSS for later down-sampling instead of FPS at every stage | PAT classification · DVS128 Gesture 11-class | system accuracy: 95.2 -> 96.0 | Helped |
| Replace channel shuffle with no shuffle (ablation) | PAT (ARPE, GSA, GN) · ModelNet40 test (256 points) | accuracy: 90.9 -> 88.8 | Hurt |
| Add ARPE instead of using MLP embedding (ablation) | PAT with GSA, channel shuffle, and group normalization · ModelNet40 test (256 points) | accuracy: 90.1 -> 90.9 | Helped |
| Replace GSA with size-matched MHA (ablation) | PAT with ARPE · ModelNet40 test (256 points) | accuracy: 90.9 -> 89.3 | Hurt |
| Replace GSA with same-hidden-size MHA (ablation) | PAT with ARPE · ModelNet40 test (256 points) | accuracy: 90.9 -> 89.9 | Hurt |
| Replace group normalization with layer normalization (ablation) | PAT with ARPE and GSA · ModelNet40 test (256 points) | accuracy: 90.9 -> 89.9 | Hurt |
| Use element-wise classification loss (ablation) | PAT classification · ModelNet40 test | accuracy | No clear effect |
| Use element-wise classification loss (ablation) | PAT classification · ModelNet40 training | convergence speed | Helped |
| Use one attention group instead of grouped attention (ablation) | PAT with GSA · ModelNet40 test (256 points) | accuracy | Hurt |
| Use eight groups rather than other tested group counts (ablation) | PAT with GSA · ModelNet40 test (256 points) | accuracy and model size trade-off | Helped |
Closest papers by meaning
- SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.86)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.85)
- On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation (NeurIPS 2025 · similarity 0.85)
- LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling (CVPR 2024 · similarity 0.85)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.85)
- GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.85)
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.