Robust metric for the evaluation of visual saliency algorithms.

Area of Science:

• Computer Vision
• Cognitive Science
• Human-Computer Interaction

Background:

• Understanding human visual attention is crucial for developing effective computational models.
• Previous studies suggest that visual search patterns can be influenced by both low-level image features and high-level cognitive strategies.

Purpose of the Study:

• To identify and quantify a general, content-independent viewing pattern in eye-tracking data.
• To investigate whether this pattern can be leveraged to distinguish between general viewing behavior and image-specific fixation patterns.
• To develop a metric for evaluating the performance of visual saliency algorithms.

Main Methods:

• Analysis of eye fixation data from 15 observers across 1003 images.
• Eigen-decomposition of the correlation matrix of fixation data to identify dominant viewing patterns.
• Correlation analysis to determine the relationship between the dominant pattern and image regions.
• Development of a robust AUC metric for evaluating saliency algorithms based on statistical analysis.

Main Results:

• A single eigenvector accounted for 23% of the eye-tracking data, indicating a consistent, content-independent viewing pattern.
• This dominant pattern was strongly correlated with the central region of the images.
• The study successfully formulated a method to differentiate between general viewing patterns and image-feature-dependent fixations.

Conclusions:

• Human visual attention exhibits a significant, repeatable pattern that is largely independent of image content, favoring the image center.
• The statistical information from the primary eigenvector can be utilized to filter general viewing patterns from image-specific ones.
• The developed AUC metric provides a robust approach for assessing the accuracy of visual saliency models.