What do saliency models predict?

Area of Science:

• Visual perception
• Computational neuroscience
• Human-computer interaction

Background:

• Saliency models predict eye movements during free viewing.
• Previous studies lacked systematic comparison across tasks using a single image set.

Purpose of the Study:

• Investigate how task differences affect saliency model prediction of human eye movements.
• Determine the optimal tasks and behavioral measures for saliency model application.

Main Methods:

• Compared three saliency models against human eye movements across four tasks (free viewing, saliency judgment, saliency search, object search).
• Utilized a novel explicit saliency judgment task with 100 observers.
• Analyzed eye movement variability across tasks and image clutter.

Main Results:

• Standard saliency models performed poorly on free viewing tasks.
• Models accurately predicted explicit saliency judgments and associated eye movements.
• Eye movement variability differed across tasks, influenced by task demands and image clutter.

Conclusions:

• Saliency models are best suited for tasks involving explicit saliency judgments.
• Task type critically influences saliency model performance and eye movement patterns.
• Understanding task-behavior relationships is key for accurate prediction of visual attention.