Continual Learning for Segment Anything Model Adaptation
- Citation Author(s):
-
Jinglong Yang
- Submitted by:
- Jinglong Yang
- Last updated:
- DOI:
- 10.21227/p0gw-6v72
23 views
- Categories:
- Keywords:
Abstract
To investigate SAM's potential in the continual scenario, we construct a benchmark for continual segmentation, called Continual SAM Adaptation Benchmark (CoSAM), which aims to systematically evaluate SAM-related algorithms's performance in the streaming scenarios. Specifically, CoSAM offers a set of 8 tasks covering diverse domains, including industrial defects, medical imaging, and camouflaged objects, to serve as a realistic and effective benchmark for evaluating current methods.
Instructions:
Datasets from different domains, serving as distinct tasks in the Continual Learning process, are organized into separate subfolders. Training and testing sets are also divided into different folders. After downloading the datasets, sequential training is conducted on the train datasets of each domain, followed by evaluation on the test datasets to assess the segmentation results.
