Research map / Point clouds / Point cloud classification
Interpretable3D: An Ad-Hoc Interpretable Classifier for 3D 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 Interpretable3D, an ad-hoc interpretable nearest-prototype classifier for 3D point clouds that estimates prototypes via optimal-transport clustering and refines them with LVQ-style updates, matching softmax backbones on ModelNet40, ScanObjectNN and ShapeNetPart.
Problem
Deep point-cloud classifiers use opaque softmax heads, and existing 3D explanations are post-hoc and can be misleading; decision-critical 3D applications need inherently interpretable models.
Approach
Replace the parametric softmax head with cosine-similarity nearest-prototype classification, where prototypes are estimated as sub-class means via Sinkhorn/APDAGD optimal transport clustering, optimized with LVQ-style reward/penalty updates, and in final epochs replaced by the most similar real training samples.
Key results
- PointNeXt + Ours reaches 94.3% OA on ModelNet40 vs 94.0% for PointNeXt; PointNet2 + Ours gains 1.0% OA over PointNet2 (93.2 vs 92.2).
- On ScanObjectNN, PointNeXt + Ours achieves 88.0% OA vs 87.7% for PointNeXt; PointMLP + Ours 85.6% vs 85.4%.
- On ShapeNetPart, PointNeXt + Ours reaches 85.6% class mIoU and 87.2% instance mIoU vs 85.2/87.0 for PointNeXt.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Interpretable3D to PointMLP | PointMLP · ModelNet40 | OA (%): 94.1 -> 94.1 | No clear effect |
| Add Interpretable3D to PointMLP | PointMLP · ModelNet40 | mAcc (%): 91.3 -> 92.0 | Helped |
| Add Interpretable3D to PointNeXt | PointNeXt · ScanObjectNN | OA (%): 87.7 -> 88.0 | Helped |
| Add Interpretable3D to PointNeXt | PointNeXt · ScanObjectNN | mAcc (%): 85.8 -> 86.5 | Helped |
| Add Interpretable3D to PointNeXt | PointNeXt · ShapeNetPart | Class mIoU: 85.2 -> 85.6 | Helped |
| Add Interpretable3D to PointNeXt | PointNeXt · ShapeNetPart | Instance mIoU: 87.0 -> 87.2 | Helped |
| Add Sinkhorn-Knopp Prototype Estimation to random prototype initialization (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 68.0±1.5 -> 80.3±0.5 | Helped |
| Use APDAGD instead of Sinkhorn-Knopp for Prototype Estimation (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 80.3±0.5 -> 81.2±0.3 | Helped |
| Add Eq. (7) Prototype Optimization to PE1 (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 80.3±0.5 -> 83.6±0.2 | Helped |
| Add Eq. (8) Prototype Optimization to PE1 (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 80.3±0.5 -> 84.1±0.1 | Helped |
| Add Eq. (7) Prototype Optimization to PE2 (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 81.2±0.3 -> 85.2±0.2 | Helped |
| Add Eq. (8) Prototype Optimization to PE2 (ablation) | PointMLP on ScanObjectNN · ScanObjectNN | OA (%): 81.2±0.3 -> 85.6±0.1 | Helped |
Surprising findings
- Prototype Optimization with Prototype Estimation (PointNeXt backbone) outperforms the softmax PointNeXt on ModelNet40 while adding interpretability.
- Its prototypes are about 600 versus 6000 for an adapted ProtoPNet, yet achieve higher OA.
Limitations
- Gains over softmax baselines are small (often under 1%) and within reported variance on several benchmarks.
- Part segmentation uses Deep Hough Voting, and results are only comparable rather than superior.
- Interpretability is evaluated qualitatively with examples and MMD/witness metrics, without human studies.
Experimental setup
Datasets: ModelNet40 ScanObjectNN ShapeNetPart
Benchmarks: ModelNet40 ScanObjectNN ShapeNetPart
Metrics: OA mAcc class mIoU instance mIoU MMD2 maximum witness value training and test throughput
Baselines: PointNet PointNet++ DGCNN PointNet2 PointMLP PointNeXt ProtoPNet
Builds on: ProtoPNet Learning Vector Quantization (LVQ2.1) MoCo momentum update Sinkhorn-Knopp optimal transport Deep Hough Voting
Closest papers by meaning
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.88)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.88)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.86)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.86)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.86)
- Novel Class Discovery for Point Cloud Segmentation via Joint Learning of Causal Representation and Reasoning (NeurIPS 2025 · 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.