Best Paper Award for Causal Egocentric Gaze Estimation Workshop Paper at GAZE'26

Jon E. Froehlich

By Jon E. Froehlich

Receicing the best paper award in front of the title slide at the workshop

Huge congrats to Jia Li and collaborators for earning the Best Paper Award at the GAZE2026 workshop at CVPR this year for our paper entitled "How Much Future Helps? A Controlled Study of Future-Privileged Supervision for Causal Egocentric Gaze Estimation."  Jia Li is a PhD student at UT-Dallas advised by Professor Yapeng Tian.

The paper contributes:

  • A controlled formulation for studying future context. We formulate egocentric gaze estimation under a causal online setting and introduce a future-privileged training framework that isolates the impact of future look-ahead while keeping the inference architecture fixed.
  • An empirical characterization of the optimal future range. Across EGTEA Gaze+ and Ego4D, we demonstrate that future-privileged supervision consistently outperforms the causal baseline, yet the benefits concentrate within a specific, bounded temporal window (1.7-3.3 seconds) rather than increasing indefinitely.
  • Practical guidance for online egocentric gaze modeling. We show that a lightweight causal decoder can effectively absorb future-aware signals during training while preserving strict causality at inference, offering actionable insights for the desi