Information-theoretic model comparison unifies saliency metrics.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Human-Computer Interaction

Background:

• Predicting human gaze (fixations) from images is crucial for understanding visual exploration and advancing computer vision.
• Existing quantitative eye movement models (saliency prediction) suffer from inconsistent evaluation metrics.
• Inconsistent results stem from differing definitions of "saliency maps" and confounding factors like central fixation bias.

Purpose of the Study:

• To develop a unified and consistent framework for evaluating image saliency prediction models.
• To resolve inconsistencies in model comparison metrics by addressing differing definitions of saliency maps.
• To disentangle model performance from image-independent factors like central bias and spatial blurring.

Main Methods:

• Framed fixation prediction models probabilistically, utilizing information gain for evaluation.
• Jointly optimized model scale, center bias, and spatial blurring within the information-theoretic framework.
• Developed pixel-level analysis to identify model failures in capturing fixation information.

Main Results:

• Re-evaluated existing metrics within the new framework yielded near-perfect agreement in model rankings.
• Successfully separated model performance from confounding factors such as center bias and spatial blurring.
• Introduced a method for detailed pixel-level analysis of model prediction errors.

Conclusions:

• The proposed information-theoretic approach provides a robust and consistent method for saliency model evaluation.
• Disentangling performance from bias factors leads to more reliable comparisons of visual attention models.
• The framework and accompanying software facilitate improved development and understanding of image-based gaze prediction.