Research map / Point clouds / Point cloud classification
Dynamic Adapter Meets Prompt Tuning: Parameter-Efficient Transfer Learning for 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 DAPT, a parameter-efficient transfer method for point cloud transformers that combines a per-token Dynamic Adapter (learned ReLU scale) with Internal Prompts generated from adapter outputs, tuning under 5% of parameters.
Problem
Full fine-tuning of pre-trained point cloud models is costly in GPU memory and storage, and existing adapter and prompt tuning methods from NLP/2D vision transfer poorly to point clouds because of static manual scales and random prompts.
Approach
A Dynamic Adapter predicts a per-token scale via ReLU of a learned linear projection, and Internal Prompts are formed by pooling the Dynamic Adapter outputs rather than adding random external prompts; a task-agnostic feature transform (TFTS) is added as well.
Key results
- DAPT on Point-BERT reaches 91.05% on ScanObjectNN OBJ_BG (+3.62 over full fine-tuning) with 1.1M tunable parameters (4.97%).
- On the hardest ScanObjectNN PB_T50_RS variant with Point-MAE, DAPT reaches 85.08% vs 85.18% for full fine-tuning and 84.94% for IDPT, with 1.1M tunable parameters.
- Compared with full fine-tuning, DAPT reduces GPU training memory by 35% at batch size 512 and tunable parameters by about 95%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DAPT | Point-BERT · ScanObjectNN PB T50 RS | overall accuracy: 83.07 -> 85.43 | Helped |
| Apply DAPT | Point-MAE · ScanObjectNN OBJ ONLY | overall accuracy: 88.29 -> 90.19 | Helped |
| Apply DAPT | RECON · ScanObjectNN OBJ ONLY | overall accuracy: 92.77 -> 92.43 | Hurt |
| Apply DAPT | Point-MAE · ScanObjectNN PB T50 RS | overall accuracy: 85.18 -> 85.08 | Hurt |
| Apply DAPT | RECON · ModelNet40 | overall accuracy: 93.0 -> 94.1 | Helped |
| Apply DAPT | Point-MAE · ShapeNetPart | Cls. mIoU: 84.19 -> 84.01 | Hurt |
| Add TFTS (ablation) | Point-MAE linear probing · ScanObjectNN PB T50 RS | overall accuracy: 75.99 -> 83.17 | Helped |
| Add Dynamic Adapter (ablation) | Point-MAE linear probing · ScanObjectNN PB T50 RS | overall accuracy: 75.99 -> 84.25 | Helped |
| Add Internal Prompt to TFTS and Dynamic Adapter (ablation) | Point-MAE · ScanObjectNN PB T50 RS | overall accuracy: 84.80 -> 84.70 | No clear effect |
| Replace Internal Prompt with External Prompt (ablation) | Point-MAE with TFTS and Dynamic Adapter · ScanObjectNN PB T50 RS | overall accuracy: 85.08 -> 84.70 | Hurt |
| Set Dynamic Adapter dimension r to 64 (ablation) | Point-MAE DAPT · ScanObjectNN PB T50 RS | overall accuracy: 84.39; 84.52; 84.18; 85.08; 84.46 | Helped |
| Use dynamic scale instead of fixed scale 1.0 (ablation) | Point-MAE DAPT · ScanObjectNN PB T50 RS | overall accuracy: 84.59 -> 85.08 | Helped |
Surprising findings
- Standard NLP/2D PETL methods such as Adapter and VPT lag full fine-tuning by 1.7% and 4.1% on the hardest ScanObjectNN variant.
- Adding external random-initialized prompts slightly hurt performance relative to no prompt tuning (-0.1%), while internal prompts helped.
Limitations
- Not evaluated on more complex tasks such as 3D object detection or generation.
- Gains on some variants are marginal or negative relative to full fine-tuning (e.g., ShapeNetPart and PB_T50_RS for Point-MAE).
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNetPart
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart
Metrics: overall accuracy mIoU tunable parameters GPU memory FLOPs
Baselines: Point-BERT Point-MAE RECON IDPT Adapter VPT-Deep LoRA AdaptFormer SSF BitFit Prefix tuning
Builds on: AdaptFormer VPT IDPT SSF Point-MAE
Closest papers by meaning
- PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning (CVPR 2025 · similarity 0.88)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.86)
- Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models (AAAI 2024 · similarity 0.85)
- Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · similarity 0.85)
- LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling (CVPR 2024 · similarity 0.85)
- MICAS: Multi-grained In-Context Adaptive Sampling for 3D Point Cloud Processing (CVPR 2025 · similarity 0.85)
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.