Gaze-estimation benchmark exposes model-collapse and transfer weaknesses

A Gdansk University of Technology study evaluated 22 deep-learning architectures on four gaze datasets. It found that careful backbone selection can outperform more elaborate methods, while cross-dataset tests exposed poor extrapolation between screen-based and wide-angle gaze distributions. Eye regions carried most predictive information.

Why it matters: Gaze systems used in accessibility, attention analysis and spatial interfaces can fail under distribution shift despite good in-dataset scores.

Impact: Developers need cross-device, demographic and domain testing, plus privacy controls for behavioural data. Scientific Reports

Geopolitics · China/United States