Instance-Level Relative Saliency Ranking With Graph Reasoning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Conventional salient object detection models struggle to differentiate the importance of multiple salient objects.
• Existing saliency ranking models have limitations in differentiating object instances or focus on sequential attention.
• A practical need exists for models that simultaneously segment salient instances and infer their relative saliency rank order.

Purpose of the Study:

• To present a novel unified model for simultaneous salient instance segmentation and relative saliency ranking.
• To address the limitations of previous salient object detection and ranking methods.
• To advance the field of salient object detection with a more comprehensive approach.

Main Methods:

• An improved Mask R-CNN is utilized for segmenting salient instances.
• A saliency ranking branch is integrated to infer relative saliency.
• A graph reasoning module combining instance interaction, contrast, and semantic priors is developed.
• A novel loss function is proposed for training the saliency ranking branch.

Main Results:

• The proposed unified model demonstrates superior effectiveness compared to previous methods.
• Experimental results validate the model's capability in simultaneous segmentation and ranking.
• The model shows practical utility in applications like adaptive image retargeting.

Conclusions:

• The developed unified model offers an effective end-to-end solution for salient instance segmentation and relative saliency ranking.
• The novel graph reasoning module and loss function contribute to improved performance.
• The proposed dataset and evaluation metric will facilitate future research in this domain.