Pluralistic Salient Object Detection.

Area of Science:

• Computer Vision
• Image Segmentation
• Machine Learning

Background:

• Conventional salient object detection (SOD) produces a single mask, failing to capture the complexity and ambiguity of real-world images.
• Defining salient objects can be subjective, varying with user intention and context.
• Existing SOD datasets often lack diverse ground truth for ambiguous saliency scenarios.

Purpose of the Study:

• Introduce pluralistic salient object detection (PSOD) to generate multiple plausible salient segmentation masks.
• Address the limitations of single-mask SOD by accounting for image complexity and subjective saliency.
• Provide new datasets and evaluation metrics for studying and advancing PSOD.

Main Methods:

• Developed two new SOD datasets: DUTS-MM with improved and multiple ground-truth masks, and DUTS-MQ with human preference scores.
• Proposed a pluralistic SOD baseline model using a Mixture-of-Experts (MOE) design with multiple prediction heads.
• Incorporated query prompts to generate diverse masks and predict human preference scores.

Main Results:

• The new DUTS-MM and DUTS-MQ datasets significantly advance the study of saliency ambiguity and human preferences.
• The proposed MOE-based PSOD framework effectively generates multiple salient object masks.
• Experimental results validate the effectiveness of the datasets and the proposed PSOD approach.

Conclusions:

• Pluralistic salient object detection is a crucial step towards more realistic image analysis.
• The introduced datasets and MOE-based framework provide a strong foundation for future PSOD research.
• This work highlights the importance of considering multiple interpretations in salient object detection.