Research map / Point clouds / Point cloud classification
PointPatchMix: Point Cloud Mixing with Patch Scoring
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PointPatchMix, a patch-level point cloud mixing augmentation whose mixed labels are weighted by content-based patch significance scores from a pre-trained Point-MAE teacher.
Problem
Mixing-based data augmentation for point cloud classification is limited: block-level and point-level mixing trade off diversity against preservation of local structure, and linear label mixing ignores that patches differ in importance.
Approach
Mix two point clouds at the patch level using EMD-based patch correspondences, and generate mixed labels by weighting each source's patch contributions with attention-based significance scores (attention weight times value norm) from a pre-trained Transformer teacher, rather than linear sample-ratio labels.
Key results
- With Point-MAE as baseline, PointPatchMix reaches 86.3% on ScanObjectNN PB-T50-RS (vs 85.1% baseline) and 94.1% on ModelNet40 with self-supervised pretraining.
- Applied to PointGPT, PointPatchMix achieves 88.3% on ScanObjectNN PB-T50-RS, up from 86.9%.
- On ModelNet40 with a from-scratch Transformer, PointPatchMix improves accuracy from 91.4% to 93.3%, versus 92.5% for PointCutMix.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PointPatchMix | Point-MAE · ScanObjectNN OBJ-ONLY | accuracy: 88.2 -> 91.0 | Helped |
| Add PointPatchMix | Point-MAE · ModelNet40 | accuracy: 93.8 -> 94.1 | Helped |
| Add PointPatchMix | Point-MAE · ModelNet40 few-shot settings | accuracy: 96.3 ± 2.5 -> 96.6 ± 2.2; 97.8 ± 1.8 -> 97.8 ± 1.7; 92.6 ± 4.1 -> 92.7 ± 3.3; 95.0 ± 3.0 -> 95.2 ± 2.7 | Helped |
| Replace block-level mixing with patch-level mixing (ablation) | Transformer on ModelNet40 · ModelNet40 | accuracy: 91.7 -> 92.4 | Helped |
| Use patch-score targets instead of linear-combination targets (ablation) | Patch-mixing Transformer · ModelNet40 | accuracy: 92.4 -> 92.9 | Helped |
| Use center points rather than all points for patch assignment (ablation) | PointPatchMix · ModelNet40 | accuracy: 92.9 -> 92.9 | No clear effect |
| Set β to 0.5 instead of 1 (ablation) | PointPatchMix · ModelNet40 | accuracy: 92.9 -> 92.8 | Hurt |
| Set β to 1.5 (ablation) | PointPatchMix · ModelNet40 | accuracy: 92.9 -> 93.3 | Helped |
| Set β to 2 instead of 1 (ablation) | PointPatchMix · ModelNet40 | accuracy: 92.9 -> 93.1 | Helped |
| Add PointPatchMix | Transformer · ModelNet40 | accuracy: 91.4 -> 93.3 | Helped |
| Add PointPatchMix instead of PointCutMix | Transformer · ModelNet40 robustness tests | accuracy: 92.5 -> 93.3; 77.8 -> 79.7; 84.2 -> 85.2; 87.2 -> 87.6; 90.0 -> 90.5; 80.6 -> 81.7; 44.8 -> 47.4; 92.1 -> 92.5 | Helped |
Surprising findings
- Block-level and point-level mixing alternatives gave smaller gains than patch-level mixing, and linear label mixing underperformed content-based patch scoring on ModelNet40 (92.4% vs 92.9%).
Limitations
- Evaluated only on point cloud classification; segmentation and detection not explored.
- Depends on a pre-trained Transformer teacher (Point-MAE) to compute significance scores.
- Gains on some settings are small (e.g., ModelNet40 with PointMLP 94.5 to 94.7).
Experimental setup
Datasets: ModelNet40 ScanObjectNN
Benchmarks: ModelNet40 ScanObjectNN
Metrics: classification accuracy few-shot accuracy (N-way K-shot, 10 runs)
Baselines: PointCutMix PointMixup RSMix PointMixSwap Point-MAE PointGPT Point-BERT PointMLP Transformer-OcCo PointNet
Builds on: PointCutMix PointMixup Point-MAE Earth Mover's Distance
Closest papers by meaning
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2024 · similarity 0.86)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 2025 · similarity 0.86)
- Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians (CVPR 2025 · similarity 0.86)
- patchDPCC: A Patchwise Deep Compression Framework for Dynamic Point Clouds (AAAI 2024 · similarity 0.85)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · 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.