Highlights
GLEVR, a new graph-based AI framework from Intel and University of Colorado researchers, achieves more than 12% improvement in recognizing fine-grained keystep actions in first-person videos and wins the Ego-Exo4D keystep recognition challenge at CVPR 2025.The system leverages multiple camera viewpoints during training but only needs a single egocentric camera during actual use, making it practical for real-world applications such as smart assistants and training systems.By representing video clips as connected nodes in a graph structure, the framework efficiently captures long-term patterns in activities while remaining computationally lightweight.