Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders

A transformer approach for estimating gaze targets that leverages the power of pretrained visual foundation models. Gaze-LLE provides a streamlined gaze architecture that learns only a lightweight gaze decoder on top of a frozen, pretrained visual encoder (DINOv2). Gaze-LLE learns 1-2 orders of magnitude fewer parameters than prior works and doesn't require any extra input modalities like depth and pose!