Research map / Point clouds / Point cloud classification
Improving Robustness of 3D Point Cloud Recognition from a Fourier Perspective
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 Frequency Adversarial Training (FAT), which uses graph-Fourier-domain high- and low-frequency adversarial examples with AdvProp BN to improve corruption robustness of point cloud classifiers on ModelNet-C and related benchmarks.
Problem
Point cloud recognition models degrade substantially under common corruptions (noise, density change, transformation), and existing corruption-robustness augmentations operate only in the spatial domain.
Approach
Analyze corruptions and model sensitivity in the graph Fourier domain (GFT of the loss Jacobian), then train with adversarial perturbations applied separately to high- and low-frequency bands, with separate BN statistics (AdvProp), and provide an OOD generalization bound based on frequency-domain adversarial robustness.
Key results
- On ModelNet-C, FAT lowers DGCNN mCE to 0.825 vs 1.000 vanilla and PCT to 0.907 vs 1.000 while keeping OA close to baseline (e.g., DGCNN OA 0.925 vs 0.926).
- Combining FAT with WOLFMix on GDANet achieves 0.537 mCE on ModelNet-C, reported as new state of the art.
- On ModelNet40-C, GDANet+WOLFMix+FAT reaches 0.555 mCE and 0.147 ERcor, improving over PCT+CutMix-R (0.635 mCE, 0.163 ERcor).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply frequency adversarial training | DGCNN · ModelNet-C | mCE: 1.000 -> 0.825 | Helped |
| Apply frequency adversarial training | PointNet · ModelNet-C | mCE: 1.422 -> 1.237 | Helped |
| Apply frequency adversarial training | PCT · ModelNet-C | mCE: 0.925 -> 0.907 | Helped |
| Apply frequency adversarial training | GDANet · ModelNet-C | mCE: 0.892 -> 0.850 | Helped |
| Combine FAT with WOLFMix | GDANet · ModelNet-C | mCE: 0.601 -> 0.537 | Helped |
| Add low-frequency adversarial samples (ablation) | FAT on PointNet · ModelNet-C | mCE: 1.306 -> 1.237 | Helped |
| Add high-frequency adversarial samples (ablation) | FAT on PointNet · ModelNet-C | mCE: 1.317 -> 1.237 | Helped |
| Divide adversarial samples into high- and low-frequency bands (ablation) | FAT on PointNet · ModelNet-C | mCE: 1.310 -> 1.237 | Helped |
| Use the AdvProp training framework (ablation) | FAT on PointNet · ModelNet-C | OA: 0.885 -> 0.902 | Helped |
| Use the AdvProp training framework (ablation) | FAT on PointNet · ModelNet-C | mCE: 1.263 -> 1.237 | Helped |
Surprising findings
- FAT w/o Advprop achieves a lower mCE than full FAT on some variants but has much worse clean accuracy.
- Clean accuracy drops are small but present (e.g., PCT drops 1.0%), the authors attribute this to an accuracy-robustness trade-off.
Limitations
- Reduces clean accuracy slightly (DGCNN 0.1%, PointNet 0.5%, PCT 1.0%, GDANet 0.6%).
- No error bars or statistical significance reported due to computational cost.
- Evaluated mainly on point cloud classification with corruption benchmarks; generalization to other tasks not shown.
Experimental setup
Datasets: ModelNet40 ModelNet-C ModelNet40-C ScanObjectNN-C
Benchmarks: ModelNet-C ModelNet40-C ScanObjectNN-C
Metrics: mCE OA mOA RmCE ERcor
Baselines: Vanilla Training Adversarial Training DUP Defense RSMix PointWOLF WOLFMix AdaptPoint CutMix-R
Builds on: AdvProp PointWOLF WOLFMix PGD AOF
Closest papers by meaning
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.87)
- Target-Guided Adversarial Point Cloud Transformer Towards Recognition Against Real-world Corruptions (NeurIPS 2024 · similarity 0.86)
- Curvature-Invariant Adversarial Attacks for 3D Point Clouds (AAAI 2024 · similarity 0.86)
- ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration (ICLR 2025 · similarity 0.85)
- Hide in Thicket: Generating Imperceptible and Rational Adversarial Perturbations on 3D Point Clouds (CVPR 2024 · similarity 0.85)
- Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack (AAAI 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.