SAM-PARSER: Fine-Tuning SAM Efficiently by Parameter Space Reconstruction
Authors: Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, Wei Shen
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that SAM-PARSER exhibits superior segmentation performance across various scenarios, while reducing the number of trainable parameters by approximately 290 times compared with current parameter-efficient fine-tuning methods. |
| Researcher Affiliation | Academia | Zelin Peng*, Zhengqin Xu*, Zhilin Zeng, Xiaokang Yang, Wei Shen Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {zelin.peng, fate311, bernardeschi, xkyang, wei.shen}@sjtu.edu.cn |
| Pseudocode | No | The paper provides mathematical formulations and schematic diagrams, but no structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code, nor does it include a link to a code repository. |
| Open Datasets | Yes | To validate the effectiveness of our proposed methods, we conduct experiments on fine-tuning SAM to five datasets. (1) CT Abdominal organ (AO) (Ma et al. 2022). (2) COCO2017 (COCO) (Lin et al. 2014). (3) PASCAL VOC2012 (VOC) (Everingham et al. 2010). (4) NWPU VHR-10 (NWPU) (Cheng et al. 2014; Cheng and Han 2016; Cheng, Zhou, and Han 2016). (5) WHU building extraction (WHU) (Ji, Wei, and Lu 2018). |
| Dataset Splits | Yes | Following (Ma and Wang 2023), we randomly split 80% medical images of AO for fine-tuning and 20% for testing. We fine-tune SAM by using natural images on the training set and evaluate the models performance on its validation set. As recommended in (Cheng, Zhou, and Han 2016), we allocate 70% remote sensing images of NWPU for fine-tuning and the remaining 30% for evaluation. |
| Hardware Specification | No | The paper describes training details such as optimizer, learning rate, weight decay, and epochs, but it does not specify any hardware details like GPU or CPU models. |
| Software Dependencies | No | The paper mentions using the Adam optimizer but does not specify any software libraries or frameworks with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | Our training employs the Adam optimizer (Kingma and Ba 2014). For medical image segmentation, the initial learning rate is set to 10^-6, and the weight decay is 5 x 10^-4 with one image per mini-batch. The number of fine-tuning epochs is set to 25. For natural and remote sensing image segmentation, we follow Sonar SAM (Wang et al. 2023a), the initial learning rate is set to 1.5 x 10^-4, and the weight decay is 5 x 10^-5 with one image per mini-batch. The number of fine-tuning epochs is set to 10. |