Research map / Point clouds / Point cloud classification
Target-Guided Adversarial Point Cloud Transformer Towards Recognition Against Real-world Corruptions
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 APCT, a point cloud transformer that adversarially drops high-significance tokens during training (via an Adversarial Significance Identifier and Target-guided Promptor) to force broader pattern use, improving robustness to real-world corruptions on ModelNet-C, ScanObjectNN-C, and ShapeNet-C.
Problem
Transformer-based point cloud recognition models overfit to localized patterns and degrade under real-world corruptions (sensor noise, dropping, adding points, etc.).
Approach
Per-stage auxiliary heads compute token significance via top-k channel-wise token indexing; significant tokens get higher dropout rates applied to attention keys per query during training, so the network mines less dominant patterns and becomes robust to local corruption.
Key results
- Achieves 72.2% mCE on ModelNet-C, reducing prior SoTA (PointGPT, 83.4%) by 11.2 points.
- Achieves 74.2% mCE on ScanObjectNN-C, 4.0 points below PointM2AE (78.2%).
- Achieves 81.4% mCE on ShapeNet-C segmentation, 10.9 points below GDANet (92.3%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adversarial dropping and auxiliary token identification (ablation) | standard point-cloud transformer · ModelNet-C | mCE: 76.2 -> 72.2 | Helped |
| add adversarial dropping and auxiliary token identification (ablation) | standard point-cloud transformer · ScanObjectNN-C | mCE: 78.8 -> 74.2 | Helped |
| use adversarial dropping without auxiliary supervision L_A (ablation) | APCT · ModelNet-C | mCE | Hurt |
| increase focal-token selection from k=1 to k=2 (ablation) | APCT · ModelNet-C | mCE: 75.1%->72.2% | Helped |
| increase focal-token selection beyond k=2 (ablation) | APCT · ModelNet-C | mCE | Mixed |
| add PointMixup augmentation | APCT · ModelNet-C | mCE: 72.2 -> 66.2 | Helped |
| add PointCutMix-R augmentation | APCT · ModelNet-C | mCE: 72.2 -> 69.7 | Helped |
| add PointWOLF augmentation | APCT · ModelNet-C | mCE: 72.2 -> 67.0 | Helped |
| add RSMix augmentation | APCT · ModelNet-C | mCE: 72.2 -> 71.3 | Helped |
| add WOLFMix augmentation | APCT · ModelNet-C | mCE: 72.2 -> 64.7 | Helped |
| replace mini-PointNet tokenization with mini-DGCNN (ablation) | APCT · ModelNet-C | mCE: 72.2 -> 71.1 | No clear effect |
| replace mini-PointNet tokenization with mini-PCT (ablation) | APCT · ModelNet-C | mCE: 72.2 -> 72.4 | No clear effect |
Surprising findings
- Adding the dropping strategy without the auxiliary identification loss increases mCE, i.e., hurts robustness.
- Increasing the selection number k beyond 2 gives marginal gains or worse results, attributed to over-mining trivial patterns.
Limitations
- Optimal utilization of the discovered patterns is not extensively explored (stated in limitations).
- Performance is sensitive to mapping ratio gamma and selection number k; asymmetric schedules degrade results.
- Mean-based comparisons on ShapeNet-C use RmCE/mCE with a single-metric emphasis; gains are not uniform across all corruption types (e.g., Drop-L and Jitter weaker than some baselines).
Experimental setup
Datasets: ModelNet40 ScanObjectNN ShapeNet
Benchmarks: ModelNet-C ScanObjectNN-C ShapeNet-C ScanObjectNN MVImageNet
Metrics: mCE mOA RmCE OA
Baselines: DGCNN PointNet PointNet++ GDANet PCT RPC PointNeXt PointM2AE PointGPT
Builds on: Point Cloud Transformer Point-MAE PointNet RPC
Closest papers by meaning
- Hide in Thicket: Generating Imperceptible and Rational Adversarial Perturbations on 3D Point Clouds (CVPR 2024 · similarity 0.86)
- Improving Robustness of 3D Point Cloud Recognition from a Fourier Perspective (NeurIPS 2024 · similarity 0.86)
- ConcreTizer: Model Inversion Attack via Occupancy Classification and Dispersion Control for 3D Point Cloud Restoration (ICLR 2025 · similarity 0.86)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.86)
- PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification (AAAI 2024 · similarity 0.85)
- Explicitly Perceiving and Preserving the Local Geometric Structures for 3D Point Cloud Attack (AAAI 2024 · similarity 0.84)
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.