Research map / Point clouds / Point cloud classification
InfoCons: Identifying Interpretable Critical Concepts in Point Clouds via Information Theory
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 InfoCons, an information-bottleneck-based attribution framework that learns an attention bottleneck with a learnable unbiased prior to extract faithful and conceptually coherent critical subsets (concepts) from point cloud classifiers.
Problem
Explaining point cloud model predictions with critical subsets that are both faithful to the model's decision and spatially coherent, since existing gradient- and pooling-based critical point methods are biased or ignore the classifier.
Approach
Applies a variational information bottleneck to point-level intermediate features: a learned attention bottleneck produces soft masks, trained with CE loss plus KL regularization against a Gaussian/uniform prior, with Gaussian noise replacing unselected features to decouple neighbor information; the resulting score map gives critical concepts.
Key results
- On DGCNN, InfoCons (20 iterations) reduces OA to 63.70% after dropping 500 critical points versus 75.08% for CP++ and 89.87% for PCSAM (1-pass), with LIME3D (1000 queries) reaching 45.22% at higher cost.
- InfoCons single-pass explanation takes about 0.01 s per point cloud, comparable to CP++, while LIME3D takes 4.54 s.
- Plugging InfoCons into SageMix raises DGCNN ModelNet40 test OA from 92.79% to 93.19%; in SI-Adv it raises ASR from 99.76% to 99.80% with lower CD/HD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace SageMix's gradient-based saliency scores with InfoCons scores | DGCNN mixup pipeline · ModelNet40 test | Test OA (%): 92.79 -> 93.19; (+0.4) | Helped |
| Rescale SI-Adv's C&W loss gradient with InfoCons scores | DGCNN SI-Adv · ModelNet40 test | ASR (%): 99.76 -> 99.80; (+0.04) | Helped |
| Rescale SI-Adv's C&W loss gradient with InfoCons scores | DGCNN SI-Adv · ModelNet40 test | Average CD (10^-1): 5.58 -> 5.47; ($-0.11$) | Helped |
| Rescale SI-Adv's C&W loss gradient with InfoCons scores | DGCNN SI-Adv · ModelNet40 test | Average HD (10^-1): 6.70 -> 6.55; ($-0.15$) | Helped |
| Use InfoCons (20it) rather than LIME3D (10^3) | DGCNN point-drop explanation · ModelNet40 test, 500 most critical points dropped | OA (%): 45.22 -> 63.70 | Hurt |
| Use InfoCons (20it) rather than InfoCons (1pass) (ablation) | DGCNN point-drop explanation · ModelNet40 test, 500 most critical points dropped | OA (%): 73.50 -> 63.70 | Helped |
| Use InfoCons (20it) rather than InfoCons (1pass) (ablation) | DGCNN point-drop explanation · ModelNet40 test, 500 most critical points dropped | Explanation time (s): 0.01 -> 0.29 | Hurt |
Surprising findings
- Gradient-based PCSAM tends to produce spatially aggregated critical points at corners across architectures, driven by a centroid shape prior.
- A point cloud of 1024 points at the same location is predicted as 'guitar' by PointNet with 99.2% certainty, indicating a biased inductive prior.
Limitations
- Optimizing the attention bottleneck is computationally heavy and some PC models have low efficiency.
- Results depend on hyperparameters beta and intermediate layer l, which are chosen manually; failure cases occur for PCT.
- Drop-accuracy metric may reward removing points that push the cloud off the data manifold rather than true faithfulness.
Experimental setup
Datasets: ModelNet40 ScanObjectNN KITTI
Benchmarks: ModelNet40 ScanObjectNN KITTI
Metrics: overall accuracy under point-drop attacks (MCD vs LCD) explanation time ASR Chamfer distance Hausdorff distance score variance
Baselines: Critical Points Critical Points++ PC Saliency Map LIME3D
Builds on: Variational Information Bottleneck Critical Points PC Saliency Map SageMix SI-Adv
Closest papers by meaning
- Interpretable3D: An Ad-Hoc Interpretable Classifier for 3D Point Clouds (AAAI 2024 · similarity 0.85)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.84)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.83)
- CausalPC: Improving the Robustness of Point Cloud Classification by Causal Effect Identification (CVPR 2024 · similarity 0.83)
- UnO: Unsupervised Occupancy Fields for Perception and Forecasting (CVPR 2024 · similarity 0.83)
- Point Cloud Matters: Rethinking the Impact of Different Observation Spaces on Robot Learning (NeurIPS 2024 · similarity 0.83)
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.