What can saliency models predict about eye movements? Spatial and sequential aspects of fixations during encoding and...

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Computer Vision

Background:

• Saliency map models predict human fixation locations but yield mixed results with natural stimuli.
• Understanding visual attention requires evaluating models against real-world scenes and behaviors.

Purpose of the Study:

• To evaluate a specific saliency map model against human eye movements during scene encoding and recognition.
• To assess the predictive power of saliency models for both individual fixation points and scanpath sequences.

Main Methods:

• Recorded participant eye movements viewing natural scene photographs during memory encoding and recognition tasks.
• Compared recorded eye movements to predictions from the Itti & Koch (2000) saliency map model.
• Analyzed fixation locations and sequences (scanpaths) for predictive accuracy.

Main Results:

• The saliency model significantly outperformed random models in predicting fixation locations at both encoding and recognition.
• Scanpath similarity across multiple viewings indicated repetitive scanning patterns.
• Saliency models alone could not fully account for observed scanpath behavior.

Conclusions:

• While saliency models are useful for predicting fixation points, they do not capture the full complexity of visual attention.
• Scanpath sequences, potentially driven by top-down or bottom-up processes, play a crucial role in visual exploration.
• Future models should integrate scanpath dynamics for more accurate predictions of visual attention.