Research map / Point clouds / Point cloud classification
GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model
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 GAPrompt, a geometry-aware PEFT method for pre-trained point cloud models that adds learnable point prompts, a Point Shift Prompter for instance-specific shape features and point shifts, and Prompt Propagation, tuning ~0.6M parameters.
Problem
Parameter-efficient fine-tuning of pre-trained 3D point cloud models: existing PEFT methods focus on token-level prompts and adapters and fail to capture the geometric information of point clouds, leaving a gap to full fine-tuning.
Approach
Adds learnable point clouds as explicit prompts alongside the input, uses a hierarchical PointNet++-style prompter to extract a global shape feature that shifts input points and enhances prompt tokens and adapters, and injects these prompts into FPS/KNN-based token propagation inside frozen transformer blocks.
Key results
- On ScanObjectNN OBJ_BG, GAPrompt with Point-GPT-L reaches 98.97% vs 97.20% for full fine-tuning, using 2.0M trainable parameters (0.55%).
- With Point-FEMAE, GAPrompt reaches 90.67% on PB_T50_RS using 0.6M trainable parameters (2.19%), surpassing full fine-tuning (90.22%).
- On ModelNet40 with Point-GPT-L, GAPrompt reaches 96.2% accuracy, versus 94.1% for full fine-tuning and 93.9% for Point-PEFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GAPrompt instead of full fine-tuning | PointGPT-L · ScanObjectNN OBJ_BG | accuracy: 97.20 -> 98.97 | Helped |
| Use GAPrompt instead of full fine-tuning | PointGPT-L · ScanObjectNN OBJ_ONLY | accuracy: 96.60 -> 96.73 | Helped |
| Use GAPrompt instead of full fine-tuning | PointGPT-L · ScanObjectNN PB_T50_RS | accuracy: 93.40 -> 94.31 | Helped |
| Use GAPrompt instead of full fine-tuning | PointGPT-L · ModelNet40 | accuracy: 94.1 -> 96.2 | Helped |
| Add Point Shift Prompter to Point Prompt (ablation) | GAPrompt with Point-FEMAE · ScanObjectNN PB_T50_RS | accuracy: 87.85 -> 89.34 | Helped |
| Add Prompt Propagation to Point Prompt and Point Shift Prompter (ablation) | GAPrompt with Point-FEMAE · ScanObjectNN PB_T50_RS | accuracy: 89.34 -> 90.67 | Helped |
| Add Prompt Enhance to Shift Head (ablation) | Point Shift Prompter in GAPrompt with Point-FEMAE · ScanObjectNN PB_T50_RS | accuracy: 88.23 -> 89.71 | Helped |
| Add Adapter Enhance to Shift Head and Prompt Enhance (ablation) | Point Shift Prompter in GAPrompt with Point-FEMAE · ScanObjectNN PB_T50_RS | accuracy: 89.71 -> 90.67 | Helped |
| Use uniform rather than cluster Gaussian initialization for point prompts (ablation) | GAPrompt · ScanObjectNN PB_T50_RS | accuracy | Helped |
| Set the number of learnable point prompts to 20 (ablation) | GAPrompt · ScanObjectNN PB_T50_RS | accuracy | Helped |
| Set adapter enhancing factor βa to 0.5 (ablation) | GAPrompt · ScanObjectNN PB_T50_RS | accuracy | Helped |
| Use permutation prompt injection after self-attention rather than the other tested settings (ablation) | Prompt Propagation in GAPrompt · ScanObjectNN PB_T50_RS | accuracy | Helped |
Surprising findings
- Point-FEMAE and ReCon full fine-tuning are surpassed by GAPrompt on OBJ_BG despite tuning under 3% of parameters.
- The ablation finds that permutation-based prompt injection after self-attention with beta_p=0.5 works best, attributed to added randomness.
Limitations
- Gains on ModelNet40 are small because the benchmark is nearly saturated.
- Some configurations underperform full fine-tuning, e.g. on OBJ_ONLY and PB_T50_RS for Point-GPT-L and ReCon variants.
Experimental setup
Datasets: ScanObjectNN ModelNet40
Benchmarks: ScanObjectNN ModelNet40
Metrics: overall accuracy trainable parameters (M) FLOPs (G)
Baselines: IDPT DAPT Point-PEFT Prefix Tuning VPT Adapter Tuning LoRA SSF AdapterFormer Linear Probing Full fine-tuning
Builds on: Point-MAE PointNet++ IDPT DAPT Point-PEFT
Closest papers by meaning
- On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation (NeurIPS 2025 · similarity 0.90)
- Point-PEFT: Parameter-Efficient Fine-Tuning for 3D Pre-trained Models (AAAI 2024 · similarity 0.90)
- Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · similarity 0.89)
- PMA: Towards Parameter-Efficient Point Cloud Understanding via Point Mamba Adapter (CVPR 2025 · similarity 0.89)
- Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding (ICML 2025 · similarity 0.88)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 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.