Predicting Human Eye Fixations via an LSTM-based Saliency Attentive Model.

Area of Science:

• Computer Vision
• Neuroscience
• Machine Learning

Background:

• Data-driven saliency prediction, utilizing Convolutional Neural Networks (CNNs), is a rapidly advancing field for gaze fixation analysis.
• Current methods often employ feed-forward networks, which may not fully capture the dynamic and focused nature of human attention.

Purpose of the Study:

• To develop a novel model for accurate saliency map prediction that incorporates neural attentive mechanisms.
• To address the inherent center bias in human eye fixation data.

Main Methods:

• A Convolutional Long Short-Term Memory (ConvLSTM) network is employed to iteratively refine saliency maps by focusing on salient image regions.
• The model learns a set of prior maps using Gaussian functions to mitigate the center bias prevalent in human gaze data.

Main Results:

• The proposed architecture demonstrates superior performance compared to current state-of-the-art methods on established saliency prediction datasets.
• Extensive evaluations confirm the robustness of individual model components across various scenarios.

Conclusions:

• The novel attentive model significantly enhances the accuracy of saliency map prediction.
• The integration of ConvLSTM and Gaussian prior maps offers a robust solution for gaze prediction challenges.