Pairwise Comparison-Based Salient Object Ranking Using Multimodal Large Models.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Salient object ranking aims to mimic human visual attention by ordering objects by importance.
• Existing methods face challenges with complex scenes, leading to inaccuracies, especially for less prominent objects.

Purpose of the Study:

• To develop a novel framework, PairwiseSOR-MLMs, for robust salient object ranking.
• To address limitations in handling object occlusion, semantic similarity, and numerous objects in complex scenes.

Main Methods:

• Utilizes object detection and instance segmentation to identify scene objects.
• Employs image inpainting to reconstruct scenes and mitigate occlusion effects.
• Leverages multimodal large models (MLMs) for pairwise object comparisons and global ranking aggregation.

Main Results:

• Achieves state-of-the-art or competitive performance on ASSR and IRSR benchmarks.
• Demonstrates significant robustness in ranking accuracy when dealing with occluded and semantically similar objects.
• The pairwise comparison approach shows promise for extending to other relative assessment tasks.

Conclusions:

• PairwiseSOR-MLMs effectively enhances salient object ranking accuracy in complex visual scenes.
• The framework's ability to handle occlusions and semantic similarities offers a significant advancement.
• The proposed pairwise comparison methodology is adaptable for broader applications in visual assessment.