Research map / Point clouds / Point cloud classification
FlatFormer: Flattened Window Attention for Efficient Point Cloud Transformer
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 |
|---|---|---|---|
| Use window-based quantization and sorting (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 60.4 -> 61.7 | Helped |
| Alternate sorting axes between blocks (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 61.1 -> 61.7 | Helped |
| Shift windows between blocks (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 61.2 -> 61.7 | Helped |
| Replace random point order with window-based sorting (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 57.8 -> 61.7 | Helped |
| Use equal-size groups with window-based sorting instead of SST equal-window partitioning | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 60.7 -> 61.7 | Helped |
| Vary window shape from 9×9 to 6×6 or 13×13 (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 61.7 -> 61.1; 61.7 -> 61.3 | No clear effect |
| Set group size to 85% of window shape instead of 50% or 125% (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 60.7 -> 61.7; 60.9 -> 61.7 | Helped |
| Increase input resolution from 0.32m to 0.28m (ablation) | FlatFormer · Waymo Open Dataset, 20% training data | Mean L2 mAPH: 61.7 -> 61.7 | No clear effect |
| Use FlashAttention for MHSA (ablation) | FlatFormer · inference latency | speedup: 1.7× | Helped |
| Pack Q, K, and V projections into one linear kernel (ablation) | FlatFormer · inference latency | speedup: 1.3× | Helped |
| Fuse FFN linear and activation layers (ablation) | FlatFormer · inference latency | speedup: 1.2× | Helped |
| Drop the final non-full group rather than mask it (ablation) | FlatFormer · inference latency | speedup: 1.1× | Helped |
Closest papers by meaning
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.91)
- LitePT: Lighter Yet Stronger Point Transformer (CVPR 2026 · similarity 0.90)
- PatchFormer: An Efficient Point Transformer With Patch Attention (CVPR 2022 · similarity 0.90)
- ConDaFormer: Disassembled Transformer with Local Structure Enhancement for 3D Point Cloud Understanding (NeurIPS 2023 · similarity 0.90)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.87)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 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.