Quantitative analysis of human-model agreement in visual saliency modeling: a comparative study.

Area of Science:

• Computer Vision
• Cognitive Science
• Robotics

Background:

• Visual attention selects relevant scene regions using top-down (task-driven) and bottom-up (stimulus-driven) factors.
• Bottom-up visual saliency modeling is crucial for computer vision and robotics, with numerous research efforts over 20 years.
• Direct comparison of existing saliency models is challenging due to varied datasets, evaluation scores, and parameter settings.

Purpose of the Study:

• To conduct an exhaustive comparison of 35 state-of-the-art visual saliency models.
• To evaluate model performance across diverse datasets including synthetic patterns, natural images, and videos.
• To establish a unified framework for assessing and advancing visual saliency modeling.

Main Methods:

• Compared 35 visual saliency models on 54 synthetic patterns, 3 natural image datasets, and 2 video datasets.
• Utilized three distinct evaluation scores for performance assessment.
• Performed computational complexity analysis and analyzed dataset biases.

Main Results:

• Identified specific models that consistently outperform others, despite variations in rankings across datasets.
• Revealed that existing datasets exhibit significant center-bias, influencing evaluation scores.
• Demonstrated that some computationally efficient models achieve competitive eye movement prediction accuracy.

Conclusions:

• The study provides a clear assessment of the current state-of-the-art in visual saliency modeling.
• It offers insights into dataset limitations and evaluation score influences.
• A unified comparison framework is proposed to guide future research and development in this field.