Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 9·5 min read
Diffusion models for eye-gaze trajectory generation using position and velocity representations
TL;DR
Two denoising diffusion probabilistic models (DDPMs) were developed to generate eye-gaze trajectories from visual-search data, achieving high fidelity in gaze-position and velocity sequences. The position model demonstrated a mean Jensen-Shannon divergence of 0.016, while the velocity model achieved 0.0065, indicating effective capture of local gaze dynamics.
✦ Why It Matters
Consider using these models to generate synthetic eye-tracking data for research without privacy concerns.
Key Takeaways
- Utilize DDPMs for generating synthetic eye-gaze data.
- Evaluate model performance using Jensen-Shannon divergence metrics.
- Explore potential applications in privacy-sensitive research areas.
- Investigate long-range gaze dynamics in future model developments.
How It Works
The models employ a FiLM-conditioned one-dimensional U-Net architecture, which allows for effective processing of gaze data by conditioning on visual features. The training involves sliding-window sequences, enabling the models to learn temporal dynamics in gaze behavior.