Research map / Point clouds / Point cloud classification
Robustness Certification for Point Cloud Models
Tobias Lorenz, Anian Ruoss, Mislav Balunović, Gagandeep Singh, Martin Vechev
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 improved max pool relaxation instead of DeepPoly (ablation) | Taylor3D on PointNet · ModelNet40 Rot_Z +-3deg, 1024 points | certified %: 28.1 -> 37.1 | Improved |
| add improved max pool relaxation instead of DeepPoly (ablation) | Taylor3D on PointNet · ModelNet40 Rot_Z +-3deg, 128 points | certified %: 72.2 -> 81.1 | Improved |
| add improved max pool relaxation (ablation) | DeepPoly on MNIST CNN · MNIST l_inf eps=0.005 | certified images %: 72.8 -> 79.0 | Improved |
| add improved max pool relaxation (ablation) | DeepPoly on MNIST CNN · MNIST l_inf eps=0.015 | certified images %: 3.7 -> 6.2 | Improved |
| add provable (IBP) training instead of natural training (ablation) | PointNet max pool · ModelNet40 l_inf eps=0.01 (64 points) | certified %: 0.0 -> 91.4 | Improved |
| add provable (IBP) training instead of natural training (ablation) | PointNet max pool · ModelNet40 l_inf eps=0.01 (64 points) | accuracy %: 85.7 -> 77.6 | Hurt |
| add adversarial (FGSM) training instead of natural training (ablation) | PointNet max pool · ModelNet40 l_inf eps=0.01 (64 points) | certified %: 0.0 -> 22.2 | Improved |
| replace max pool with average pool (ablation) | PointNet natural training · ModelNet40 (64 points) | accuracy %: 85.7 -> 81.9 | Hurt |
| replace max pool with average pool (ablation) | PointNet adversarial training · ModelNet40 l_inf eps=0.01 (64 points) | certified %: 22.2 -> 47.7 | Improved |
| increase max pool group size from 4 to 12 (ablation) | DeepPoly improved max pool · ModelNet40 Rot_Z +-3deg | certified %: 93.5 -> 93.5 | No clear effect |
| use Taylor3D instead of DeepG3D | 3DCertify · Twist o Taper o Rot_Z (10, 0.1, 1deg) | running time ms: 65447ms -> 65ms (1000x speed-up) | Improved |
| use DeepG3D instead of Taylor3D | 3DCertify · Twist o Rot_Z (20, 1deg) | certified %: 7.6 -> 16.3 | Improved |
Closest papers by meaning
- Certified L2-Norm Robustness of 3D Point Cloud Recognition in the Frequency Domain (AAAI 2026 · similarity 0.90)
- PointCert: Point Cloud Classification With Deterministic Certified Robustness Guarantees (CVPR 2023 · similarity 0.90)
- TPC: Transformation-Specific Smoothing for Point Cloud Models (ICML 2022 · similarity 0.90)
- SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.87)
- PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion (NeurIPS 2025 · similarity 0.86)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (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.
