Personalized Saliency and Its Prediction.

Area of Science:

• Computer Vision
• Cognitive Psychology
• Machine Learning

Background:

• Existing visual saliency models predict a single map for all observers, neglecting individual attention variations.
• Psychology research indicates observer attention differs, especially with complex scenes containing multiple salient objects.

Purpose of the Study:

• To investigate heterogeneous visual attention patterns across observers.
• To develop models that predict personalized saliency maps (PSM) by accounting for individual preferences and image content.

Main Methods:

• Constructed a novel personalized saliency dataset.
• Decomposed personalized saliency maps into universal saliency maps (USM) and user-specific discrepancy maps.
• Proposed a multi-task convolutional neural network (CNN) and a CNN with Person-specific Information Encoded Filters (CNN-PIEF) to predict discrepancy maps.

Main Results:

• Demonstrated the effectiveness of the proposed CNN and CNN-PIEF models in predicting personalized saliency maps.
• Showcased the generalization capability of the models for predicting saliency in unseen observers.

Conclusions:

• The developed models successfully predict personalized saliency maps, capturing individual differences in visual attention.
• This work advances saliency modeling by incorporating observer-specific factors, moving beyond universal predictions.