Weakly supervised salient object detection via bounding-box annotation and SAM model

• DOI: 10.3934/era.2024074
• Publication Date: 2024-02-20T09:56:28Z
• Authors: Xiangquan Liu; Xiaoming Huang
• Keywords: Minimum bounding box; Bounding overwatch; Supervised Learning

<abstract><p>Salient object detection (SOD) aims to detect the most attractive region in an image. Fully supervised SOD based on deep learning usually needs a large amount of data with human annotation. Researchers have gradually focused task using weakly annotation such as category, scribble, and bounding-box, while these existing methods achieve limited performance demonstrate huge gap fully methods. In this work, we proposed one novel two-stage method bounding-box recent visual model Segment Anything (SAM). first stage, regarded box prompt SAM generate initial labels completeness check inversion exclude low quality labels, then selected reliable pseudo for training model. second used predict saliency map excluded images adopted everything mode segmentation candidates, fused candidates labels. Finally all generated two stages train refined We also designed simple but effective model, which can capture rich global context information. Performance evaluation four public datasets showed that significantly outperforms other achieves comparable methods.</p></abstract>