Visual Scanpath Prediction Using IOR-ROI Recurrent Mixture Density Network.

Area of Science:

• Computer Vision
• Cognitive Science
• Artificial Intelligence

Background:

• Visual scanpaths, representing eye movements, are crucial for understanding visual attention and search behavior.
• Predicting these scanpaths aids in modeling how humans process visual information.

Purpose of the Study:

• To develop a computational model for generating human-like visual scanpaths under task-free conditions.
• To simultaneously predict the sequence of fixation positions and their durations.

Main Methods:

• An 'Inhibition of Return - Region of Interest' (IOR-ROI) recurrent mixture density network framework was developed.
• The model integrates bottom-up and semantic features using Convolutional Neural Networks (CNNs).
• A dual Long Short-Term Memory (LSTM) unit (IOR-LSTM and ROI-LSTM) captures Inhibition of Return dynamics and gaze shift behavior, with fixation duration predicted by a regression network.

Main Results:

• The proposed IOR-ROI framework demonstrated superior performance in predicting visual scanpaths on the OSIE and MIT datasets.
• The model successfully integrated visual features and captured complex eye movement dynamics.
• Simultaneous prediction of fixation positions and durations was achieved, reflecting human-like scanpath generation.

Conclusions:

• The developed IOR-ROI model offers a significant advancement in predicting visual scanpaths.
• This approach provides a robust method for modeling overt human visual attention and search behavior.
• The framework's ability to handle individual variations in eye movement patterns is a key strength.