Enhancing Object Segmentation via Few-Shot Learning with Limited Annotated Data

Significant advancements in machine learning in recent years have revolutionized multiple sectors. The Segment-Anything Model (SAM) is a notable example of state-of-the-art image segmentation. Despite claims of zero-shot generalization, SAM exhibits limitations in specific scenarios like medical mammography images. SAM generates three segmentation masks per image to address this and recommends selecting the one with the highest confidence score. However, this is not always the optimal choice. This paper introduces a system that extends SAM’s segmentation capabilities by automatically selecting the correct mask, leveraging few-shot learning methods and an Out-of-Distribution threshold strategy. Several backbones were subjected to experimentation, highlighting the relationship between the support set size and the model’s accuracy.