GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model
Authors: Zixiang Ai, Zichen Liu, Yuanhang Lei, Zhenyu Cui, Xu Zou, Jiahuan Zhou
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that GAPrompt significantly outperforms state-of-the-art PEFT methods and achieves competitive results compared to full fine-tuning on various benchmarks, while utilizing only 2.19% of trainable parameters. Our code is available at https: //github.com/zhoujiahuan1991/ ICML2025-GAPrompt. |
| Researcher Affiliation | Academia | 1Wangxuan Institute of Computer Technology, Peking University, Beijing, China 2State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou, China 3School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. Correspondence to: Jiahuan Zhou <EMAIL>. |
| Pseudocode | No | The paper describes methods with mathematical formulations (e.g., in Section 3 and Appendix B) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https: //github.com/zhoujiahuan1991/ ICML2025-GAPrompt. |
| Open Datasets | Yes | Scan Object NN. The Scan Object NN (Uy et al., 2019) is a highly challenging 3D dataset comprising 15K real-world objects across 15 categories. [...] Model Net40. Model Net40 (Wu et al., 2015) comprises 12,311 pristine 3D CAD models across 40 categories, with complete, uniform, and noise-free point clouds that simplify the task. |
| Dataset Splits | Yes | Scan Object NN. The Scan Object NN (Uy et al., 2019) is a highly challenging 3D dataset comprising 15K real-world objects across 15 categories. [...] Model Net40. Model Net40 (Wu et al., 2015) comprises 12,311 pristine 3D CAD models across 40 categories, with complete, uniform, and noise-free point clouds that simplify the task. Following baselines, we sample 1024 points per instance. |
| Hardware Specification | Yes | All experiments are conducted on a single Ge Force RTX 4090 using Py Torch version 1.13.1. |
| Software Dependencies | Yes | All experiments are conducted on a single Ge Force RTX 4090 using Py Torch version 1.13.1. |
| Experiment Setup | Yes | We adopt downstream fine-tuning configurations in alignment with the pioneering work Point-MAE (Pang et al., 2022). The detailed configurations are provided in Table 5. For example, when fine-tuning on Scan Object NN (Uy et al., 2019), the training process spans 400 epochs, using a cosine learning rate scheduler (Loshchilov & Hutter, 2022) that starts at 5e-4, with a 10-epoch warm-up period. The Adam W optimizer (Loshchilov & Hutter, 2019) is employed. [...] Table 5. Training details for downstream fine-tuning. Dataset Scan Object NN Model Net Optimizer Adam W Learning rate 5e-4 Weight decay 5e-2 Learning rate scheduler cosine Training epochs 400 Warmup epochs 10 Batch size 32 Point Prompt number 20 Prompt enhancing factor 0.5 Adapter enhancing factor 0.5 Number of points 2048 Number of point patches 128 Point patch size 32 |