EgoLife & EgoGPT

Toward an egocentric life assistant (CVPR 2025): six participants lived together for a week wearing AI glasses, producing the EgoLife dataset — 300 hours of egocentric, interpersonal, multiview, multimodal daily-life recordings with intensive annotation — plus the EgoLifeQA long-context QA suite. The companion EgoGPT-7B is an omni-modal model tuned for egocentric video/audio understanding, with EgoRAG for ultra-long-context memory and retrieval.

NTU-led (Ziwei Liu) through the LMMs-Lab community. EgoGPT builds on open LMM backbones, not a from-scratch pretrain; the line continues in later EgoLife releases.