Improving Human Visual Attention Prediction Using Deep Learning and Webcam Eye Tracking

Overview

To evaluate advertising effectiveness, companies often rely on 3M’s Visual Attention Software (VAS). 3M VAS is a gaze prediction tool intended to provide advertisers with vital information, predicting what people focus on when gazing at advertisements. Recent research suggests 3M VAS may lack precision due to its training methods and data integration. Furthermore, 3M VAS is limited to static images. I believed my custom Artificial Intelligence Model would be more accurate in predicting viewer gaze attention due to the way my model was made and trained. I engineered a custom saliency model using a ResNet50 encoder-decoder architecture in Python, optimized for Apple’s M4 silicon chip. My model was trained over 50 epochs and more than 52,000 gaze points, using an AdamW optimizer and a custom Gaussian heatmap generator to simulate human foveal vision. By training on real-time human behavior rather than synthetic datasets, this model provides a more reliable, high-fidelity tool for predictive marketing and video analytics.

Competition history

• CSEF 2026 Cognitive Science (Junior Division) · Entry J-06-06

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Behavioural and Social Sciences projects
• CSEF Behavioural and Social Sciences projects
• Every CSEF 2026 project