Research map / Point clouds / Point cloud classification
Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis
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 Point-PRC, a plug-and-play regulation framework (mutual agreement, text diversity, and model ensemble constraints) for prompt tuning large 3D models, improving 3D domain generalization while also boosting task-specific accuracy, and introduces three new 3DDG benchmarks.
Problem
Prompt tuning large 3D multi-modal models (ULIP, ULIP-2, PointCLIP) boosts seen-class accuracy but severely degrades generalization to unseen classes and domains, falling below zero-shot performance.
Approach
Constrain learnable prompts to stay in agreement with frozen foundation-model features and predictions (L1 feature alignment plus KL prediction alignment), diversify text descriptions from LLMs and templates, and ensemble prompt parameters across epochs with Gaussian weights.
Key results
- Base-to-new: ULIP-2+RC raises average harmonic mean from 72.57% to 79.48% over 5 datasets; ULIP+RC improves new-class accuracy from 49.01% to 61.93%.
- Cross-dataset OOD: ULIP-2+RC averages 51.19% vs 44.99% for ULIP-2 across five target domains.
- Few-shot and PointDA: ULIP-2+RC reaches 74.74% average on PointDA vs 64.84% for ULIP-2 (16-shot), and 1-shot ULIP-2 on ModelNet40 reaches 73.63%, above zero-shot 71.23%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Point-PRC regulation constraints | ULIP-2 prompt tuning · Base-to-new (5-dataset average) | HM: 72.57 -> 79.48 | Helped |
| add Point-PRC regulation constraints | ULIP prompt tuning · Base-to-new (5-dataset average) | new accuracy: 49.01 -> 61.93 | Helped |
| add Point-PRC regulation constraints | ULIP-2 prompt tuning · Cross-dataset OOD (5-target average) | accuracy: +6.20% | Helped |
| add Point-PRC regulation constraints | ULIP-2 prompt tuning · ModelNet-C (corruption severity 2, 7-type average) | accuracy: +4.78% | Helped |
| add Point-PRC regulation constraints | ULIP-2 prompt tuning · PointDA (6-setting average) | accuracy: 64.84(4.04) -> 74.74(2.52) | Helped |
| add mutual agreement constraint only (ablation) | ULIP-2 · Base-to-new (5-dataset average) | HM: 72.57 -> 73.18 | Helped |
| add text diversity constraint only (ablation) | ULIP-2 · Base-to-new (5-dataset average) | HM: 72.57 -> 74.53 | Helped |
| add model ensemble constraint only (ablation) | ULIP-2 · Base-to-new (5-dataset average) | HM: 72.57 -> 78.89 | Helped |
| add mutual agreement and text diversity constraints, without MEC (ablation) | ULIP-2 · Base-to-new (5-dataset average) | new accuracy: 67.91 -> 75.59 | Helped |
| use L1 rather than cosine distance in MAC (ablation) | ULIP-2 + MAC + TDC + MEC · Base-to-new (5-dataset average) | HM: 76.10 -> 79.48 | Helped |
| use L1 rather than MSE distance in MAC (ablation) | ULIP-2 + MAC + TDC + MEC · Base-to-new (5-dataset average) | HM: 77.97 -> 79.48 | Helped |
| use manual point-cloud descriptions rather than GPT-3.5 descriptions (ablation) | ULIP-2 + Point-PRC · Base-to-new (5-dataset average) | HM: 76.92 -> 79.48 | Helped |
Surprising findings
- Regulation constraints improve base (seen) class accuracy as well as new-class accuracy, not trading one for the other.
- LLM-generated descriptions (GPT-3.5, GPT-4) underperform 64 manual templates on new-class accuracy, while PointLLM descriptions perform better among LLM sources.
Limitations
- Validated only on object-level point cloud recognition, not scene-level segmentation or detection.
- Text descriptions are fused by simple average pooling; more sophisticated fusion is unexplored.
- Omni3D cross-dataset gains are very limited, with PointCLIP-series and ULIP reaching under 10% accuracy.
Experimental setup
Datasets: ModelNet40 ScanObjectNN ShapeNetCoreV2 ModelNet-C PointDA Sim-to-Real Omni3D Objaverse-LVIS
Benchmarks: ModelNet40 ScanObjectNN ShapeNetCoreV2 ModelNet-C PointDA Sim-to-Real Omni3D Objaverse-LVIS
Metrics: accuracy harmonic mean of base and new accuracy
Baselines: PointCLIP PointCLIP V2 ULIP ULIP-2 MetaSets PDG PointDAN I-OODG
Builds on: PointCLIP PointCLIP V2 ULIP ULIP-2 PPT MetaSets PDG
Closest papers by meaning
- GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.89)
- Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training (CVPR 2024 · similarity 0.89)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.87)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.87)
- PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model (AAAI 2025 · similarity 0.87)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.87)
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.