Semi-Supervised Gaze Estimation via Disentangled Subspace Contrastive Learning

Researchers introduced DSCL, a plug-and-play semi-supervised gaze estimation framework from Qida Tan and Wenchao Du.

DSCL disentangles gaze features into pitch- and yaw-specific subspaces using Jacobian regularization and contrastive ranking on labeled and unlabeled data.

The method achieves competitive in-domain and cross-domain results with only 5% to 20% of annotated training data.

By reducing reliance on costly labels while preserving accuracy, DSCL could make gaze tracking easier to deploy in real-world settings.