UNIEGO: Proxies as Mediators for Unified Egocentric Video Representation Learning

Wenhao Chi, Arkaprava Sinha, Dominick Reilly, Hieu Le, Srijan Das|June 18, 2026arXiv

Key Takeaway

Using proxy models as intermediaries between diverse teachers prevents conflicting gradients and enables learning richer egocentric representations from heterogeneous knowledge sources—achieving better results than naive multi-teacher distillation.

Summary

This paper introduces UNIEGO, a unified egocentric video encoder trained through a novel multi-teacher distillation framework.

multimodal training architecture

Key Terms

knowledge-distillation multi-teacher-learning egocentric-perception proxy-model selective-distillation