Global-and-Local Collaborative Learning for Co-Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Co-salient object detection (CoSOD) aims to identify common salient objects across a group of related images.
• Effective extraction of inter-image correspondence is critical for advancing CoSOD performance.
• Existing methods often struggle with comprehensive inter-image relationship modeling.

Purpose of the Study:

• To propose a novel architecture, GLNet, for enhanced co-salient object detection.
• To effectively capture both global and local inter-image correspondences for improved CoSOD.
• To demonstrate the efficacy of GLNet against state-of-the-art methods using benchmark datasets.

Main Methods:

• Introduced a global-and-local collaborative learning (GLNet) architecture.
• Employed 3-D convolution for global correspondence modeling (GCM) by treating images as time slices.
• Utilized pairwise correlation transformation (PCT) for local correspondence modeling (LCM) and integrated them via a global-and-local correspondence aggregation (GLA) module.
• Implemented an intra-and-inter weighting fusion (AEWF) module for adaptive feature integration.

Main Results:

• GLNet successfully extracts comprehensive inter-image relationships from both global and local perspectives.
• The model demonstrates superior performance in predicting co-saliency maps.
• Achieved state-of-the-art results on three prevailing CoSOD benchmark datasets.

Conclusions:

• GLNet offers a significant advancement in co-salient object detection by effectively modeling inter-image correspondences.
• The proposed architecture achieves superior performance compared to 11 competitors, even when trained on a smaller dataset.
• GLNet's collaborative learning approach provides a robust framework for future CoSOD research.