Research map / Point clouds / Point cloud classification
Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models
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-PEFT, a parameter-efficient fine-tuning framework for 3D point cloud pre-trained transformers that combines a memory-bank-based point-prior prompt with a geometry-aware local adapter, training about 5% of parameters.
Problem
Parameter-efficient fine-tuning has been developed for language and 2D models, but a specialized PEFT method for 3D point cloud pre-trained models remains under-explored, and full fine-tuning is expensive.
Approach
Freeze the pre-trained 3D transformer and add two modules: a Point-prior Prompt whose tokens are enhanced by attention over a training-set feature bank (training-free cache), and a Geometry-aware Adapter that performs FPS/kNN local aggregation with shared self-attention to capture fine-grained local geometry.
Key results
- On ScanObjectNN PB-T50-RS, Point-PEFT with Point-MAE (dagger, stronger augmentation) reaches 89.1% accuracy with 0.7M trainable parameters vs 88.1% for full fine-tuning with 22.1M parameters (+1.0%).
- On ModelNet40, Point-PEFT reaches 94.2% with Point-MAE, +1.0% over full fine-tuning, using under 4% of learnable parameters.
- Point-PEFT outperforms Prompt Tuning by +5.5%, Adapters by +2.4%, LoRA by +2.8%, and Bias Tuning by +4.1% on the same ScanObjectNN split (Point-MAE dagger).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Point-BERT with Point-PEFT instead of full fine-tuning | Point-BERT · ScanObjectNN PB-T50-RS | accuracy: +1.9% | Helped |
| fine-tune Point-M2AE with Point-PEFT instead of full fine-tuning | Point-M2AE · ScanObjectNN PB-T50-RS | accuracy: +0.0 | No clear effect |
| fine-tune Point-M2AE with Point-PEFT instead of full fine-tuning using stronger augmentation | Point-M2AE · ScanObjectNN PB-T50-RS | accuracy: +0.1% | Helped |
| fine-tune Point-MAE with Point-PEFT instead of full fine-tuning | Point-MAE · ScanObjectNN PB-T50-RS | accuracy: +0.3% | Helped |
| fine-tune Point-MAE with Point-PEFT instead of full fine-tuning using stronger augmentation | Point-MAE · ScanObjectNN PB-T50-RS | accuracy: +1.0% | Helped |
| add both Point-prior Prompt and Geometry-aware Adapter to the baseline (ablation) | Point-MAE† · ScanObjectNN PB-T50-RS | accuracy: +0.5% | Helped |
| add Bias Tuning to Point-prior Prompt and Geometry-aware Adapter (ablation) | Point-MAE† with Point-PEFT · ScanObjectNN PB-T50-RS | accuracy: +0.5% | Helped |
| add the point-prior bank to learnable prompt tokens (ablation) | Point-prior Prompt · ScanObjectNN PB-T50-RS | accuracy: +0.5% | Helped |
| add shared learnable coordinates to prompt tokens (ablation) | Point-prior Prompt · ScanObjectNN PB-T50-RS | accuracy: +0.7% | Helped |
| add local-aggregation operations (FPS, k-NN, pooling, and propagation) (ablation) | Geometry-aware Adapter · ScanObjectNN PB-T50-RS | accuracy: +1.2% | Helped |
| add intra-group self-attention (ablation) | Geometry-aware Adapter · ScanObjectNN PB-T50-RS | accuracy: +0.5% | Helped |
| add the final bottleneck-layer MLP (ablation) | Geometry-aware Adapter · ScanObjectNN PB-T50-RS | accuracy: +0.4% | Helped |
Surprising findings
- Inserting prompt tokens into all blocks does not yield significant improvement; prompting earlier blocks is preferable.
- On Point-M2AE, the gain over full fine-tuning on ScanObjectNN is negligible (+0.0 without stronger augmentation, +0.1 with it).
Limitations
- Gains over full fine-tuning are small for some backbones (e.g., +0.0 to +0.3 for some settings).
- Prompt length, depth, and adapter hyperparameters are tuned per backbone, with no unified setting.
- Evaluation limited to 3D shape classification and part segmentation on three pre-trained point cloud models.
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNetPart
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart
Metrics: accuracy mIoU_C mIoU_I number of trainable parameters
Baselines: Full fine-tuning Prompt Tuning Adapters LoRA Bias Tuning IDPT Point-NN PointNet PointNet++ PointMLP Point-PN
Builds on: Point-MAE Point-BERT Point-M2AE PointNet++ AdamW
Closest papers by meaning
- On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation (NeurIPS 2025 · similarity 0.93)
- GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model (ICML 2025 · similarity 0.90)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.87)
- MoST: Efficient Monarch Sparse Tuning for 3D Representation Learning (CVPR 2025 · similarity 0.87)
- Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · 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.