Research map / Point clouds / Point cloud classification
SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds
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 SMART-PC, a skeleton-prediction pre-training framework for test-time training of 3D point cloud classifiers that adapts by updating only BatchNorm statistics (no backpropagation) for real-time speed.
Problem
Test-time training for 3D point cloud classification under corruption shifts is slow because existing methods such as MATE rely on backpropagation during adaptation, and their reconstruction-based features are sensitive to surface noise.
Approach
Pre-train a Point-MAE-style network jointly with a patch-level skeletal-sphere prediction branch (point-to-sphere, sampling and radius losses) so features are robust to corruption; at test time, adapt only BatchNorm running statistics without backpropagation, or update all parameters in a backprop mode.
Key results
- Online adaptation with all parameters reaches 72.9% on ModelNet40-C vs 69.6% for MATE-Online, and 47.4% on ScanObjectNN-C vs 36.9%.
- Backpropagation-free SMART-PC-Online† reaches 70.8% on ModelNet40-C and 46.7% on ScanObjectNN-C, and about 59.5 FPS on ModelNet40-C vs 10.8 FPS for MATE.
- Source-only SMART-PC-SO averages 61.7% on ModelNet40-C vs 53.7% for MATE-SO, and 64.5% on ShapeNet-C vs 56.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use skeleton pretraining instead of reconstruction pretraining | SMART-PC source-only model · ModelNet40-C | top-1 accuracy (%): 54.4 -> 61.7 | Helped |
| use skeleton pretraining instead of reconstruction pretraining | SMART-PC source-only model · ScanObjectNN-C | top-1 accuracy (%): 34.5 -> 38.7 | Helped |
| use skeleton pretraining instead of reconstruction pretraining | SMART-PC source-only model · ShapeNet-C | top-1 accuracy (%): 56.5 -> 64.5 | Helped |
| use SMART-PC online adaptation with skeletal loss and backpropagation instead of MATE-Online | Point-MAE · ModelNet40-C | top-1 accuracy (%): 70.6 -> 72.9 | Helped |
| use SMART-PC online adaptation with skeletal loss and backpropagation instead of MATE-Online | Point-MAE · ScanObjectNN-C | top-1 accuracy (%): 36.9 -> 47.4 | Helped |
| use SMART-PC online adaptation with skeletal loss and backpropagation instead of MATE-Online | Point-MAE · ShapeNet-C | top-1 accuracy (%): 69.1 -> 67.1 | Hurt |
| update only BatchNorm statistics instead of adapting with skeletal loss and backpropagation | SMART-PC online adaptation · ModelNet40-C | top-1 accuracy (%): 72.9 -> 70.8 | Hurt |
| update only BatchNorm statistics instead of adapting with skeletal loss and backpropagation | SMART-PC online adaptation · ScanObjectNN-C | top-1 accuracy (%): 47.4 -> 46.7 | Hurt |
| sum encoder and decoder features for classification (ablation) | SMART-PC · ModelNet-C standard mode | accuracy (%): 62.4 -> 63.1 | Helped |
| use random rotations instead of no augmentation during adaptation (ablation) | SMART-PC online adaptation · ScanObjectNN-C | accuracy (%): 46.42 -> 47.36 | Helped |
| remove radius regularization loss (ablation) | SMART-PC skeleton pretraining · ModelNet40-C online adaptation | corrupted accuracy (%): 72.84 -> 67.82 | Hurt |
| remove point-to-sphere loss (ablation) | SMART-PC skeleton pretraining · ModelNet40-C online adaptation | corrupted accuracy (%): 72.84 -> 67.79 | Hurt |
Surprising findings
- On ShapeNet-C, MATE-Online (69.1%) slightly beats SMART-PC-Online (67.1%), and SMART-PC-Online is worse than its backpropagation-free variant on ShapeNet-C (65.9% vs 67.1% is the reverse; SMART-PC-Online is higher), so the gains are not uniform across datasets.
Limitations
- Standard and full-parameter online modes require backpropagation, so only the BatchNorm-only variant is fast.
- On ShapeNet-C, SMART-PC-Online does not beat MATE-Online (67.1% vs 69.1%).
- Performance on some corruptions such as background and LiDAR remains low (e.g., ~10-12% for several ScanObjectNN-C and ShapeNet-C corruptions).
Experimental setup
Datasets: ModelNet40 ShapeNetCore-v2 ScanObjectNN ModelNet40-C ShapeNet-C ScanObjectNN-C
Benchmarks: ModelNet40-C ShapeNet-C ScanObjectNN-C
Metrics: top-1 classification accuracy frames per second
Baselines: MATE Org-SO DUA TTT-Rot SHOT T3A TENT CloudFixer DDA SVWA BFTT3D
Builds on: MATE Point-MAE Point2Skeleton BFTT3D TENT
Closest papers by meaning
- PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion (NeurIPS 2025 · similarity 0.88)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.87)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 2025 · similarity 0.86)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (NeurIPS 2025 · similarity 0.86)
- MIRACLE 3D: Memory-efficient Integrated Robust Approach for Continual Learning on 3D Point Clouds via Shape Model Construction (ICLR 2025 · similarity 0.86)
- SPEAL: Skeletal Prior Embedded Attention Learning for Cross-Source Point Cloud Registration (AAAI 2024 · similarity 0.86)
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.