Predicting Human Saccadic Scanpaths Based on Iterative Representation Learning.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Cognitive Science

Background:

• Visual attention is crucial for information acquisition during scene exploration.
• Existing attention models primarily focus on static salient locations, neglecting dynamic saccadic eye movements.
• Saccadic scanpath prediction remains an underexplored area in computational attention modeling.

Purpose of the Study:

• To develop an iterative representation learning framework for saccadic scanpath prediction.
• To model the dynamic process of visual attention, incorporating saccade characteristics.
• To improve the accuracy of predicting human eye movements during visual search.

Main Methods:

• Proposed an iterative framework where saccades predict and update visual representations.
• Introduced a Bayesian definition of saccade, combining perceptual residuals (autoencoder error) and spatial location (saccade amplitude, center-weighted mechanism).
• Implemented an updating phase by retraining the network with samples around the current fixation.

Main Results:

• The model successfully replicates fundamental psychophysical properties observed in visual search tasks.
• Achieved superior performance compared to existing methods on multiple benchmark eye-tracking datasets.
• Demonstrated the effectiveness of incorporating dynamic saccade information into attention models.

Conclusions:

• The proposed iterative representation learning framework offers a more comprehensive approach to visual attention modeling.
• Integrating dynamic saccade prediction enhances the understanding and simulation of human visual exploration.
• This work advances the field of computational attention by addressing the limitations of static saliency models.