Re-Thinking Co-Salient Object Detection.

Area of Science:

• Computer Vision and Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Co-salient object detection (CoSOD) extends salient object detection (SOD) to identify co-occurring objects in image groups.
• Existing CoSOD datasets suffer from data bias, assuming visual similarity, which limits real-world applicability where semantic similarity is key.

Purpose of the Study:

• To introduce a more challenging benchmark, CoSOD3k, to overcome data bias in CoSOD datasets.
• To propose a unified, trainable CoSOD framework addressing the lack of integrated approaches in the field.
• To provide a comprehensive analysis of existing CoSOD algorithms and establish a performance baseline.

Main Methods:

• Development of the CoSOD3k benchmark with 3,316 images across 160 groups, featuring hierarchical annotations and diverse visual characteristics.
• Introduction of CoEG-Net, a novel framework integrating co-attention for efficient common information learning, augmenting the EGNet model.
• Benchmarking of 18 state-of-the-art algorithms across three datasets (iCoSeg, CoSal2015, CoSOD3k) with detailed group-level performance analysis.

Main Results:

• CoSOD3k provides a more realistic and challenging dataset for evaluating CoSOD models.
• CoEG-Net demonstrates improved model scalability and stability by leveraging prior SOD datasets and a co-attention strategy.
• Comprehensive benchmarking reveals performance variations across algorithms and datasets, highlighting areas for future improvement.

Conclusions:

• The proposed CoSOD3k benchmark and CoEG-Net framework offer significant advancements for the CoSOD field.
• Addressing data bias and developing unified frameworks are crucial for practical CoSOD applications.
• Further research is needed to tackle remaining challenges in co-salient object detection.