Segment Anything Model for detecting salient objects with accurate prompting and Ladder Directional Perception

Abstract

Salient object detection (SOD) focuses on finding, mining, and locating the most salient objects in an image. In recent years, with the introduction of SAM, image segmentation models have gradually become more unified. However, applying SAM to SOD still requires further exploration and effort. SOD relies on the extraction of multi-scale information. To enable SAM to perceive and adapt to multi-scale features, we propose the Cross-resolution Modeling Adapter, which is designed to encode the global information of features at different scales while achieving unified modeling of cross-resolution semantics. To aid the fusion of multi-scale features, we introduce the Ladder Directional Perception Fusion Module, which not only broadens the available feature space but also perceives and encodes the long-term and short-term dependencies in a stepped manner. Extensive experiments have demonstrated the effectiveness of the proposed method.