Our latest paper “A Domain Adaptive Deep Learning Solution for Scanpath Prediction of Paintings” is now available on arXiv!
The Problem
One of the key challenges in visual attention research is the domain gap: models trained on natural images (photos) don’t perform well on paintings, and collecting eye-tracking data for every type of image is expensive and time-consuming.
Our Solution
We developed a domain adaptation approach that allows models trained on natural images to work effectively on paintings, without requiring extensive labeled data for the target domain.
Key Contributions
- A domain adaptation framework specifically designed for scanpath prediction
- Techniques to bridge the gap between natural images and artistic content
- Comprehensive evaluation showing improved performance on painting datasets
Why This Matters
This work is important for:
- Cultural Heritage: Making visual attention models accessible for museum and art applications
- Data Efficiency: Reducing the need for expensive eye-tracking data collection
- Generalization: Building models that work across different visual domains
Links
This paper was also presented at CBMI 2022. The combination of self-supervised learning (our CVPR workshop paper) and domain adaptation provides a powerful toolkit for scanpath prediction in data-scarce scenarios.