CVPR 2026 is the world’s leading conference in computer vision and artificial intelligence, the ultimate stage where breakthroughs shaping the future of the field are unveiled. Within this prestigious framework, held from June 3–7, 2026, in Denver, Colorado, Enmin Zhong and Carlos Roberto del Blanco Adán presented their innovative paper: "Hand Trajectory Fusion for Egocentric Natural Language Query Grounding."

Their research addresses a key challenge in first-person (egocentric) videos: localizing an exact moment using a natural language text query. While traditional methods only focus on video appearance, overlooking the fact that 41% of actions happen during hand-object manipulation, they introduce a hand-trajectory encoder. This system translates hand movements into highly semantic kinematic features and fuses them with video-text data using an adaptive cross-attention strategy.

Results on the demanding Ego4D dataset prove that hands are the missing piece, delivering significant improvements of up to +4.32 in quantity/state queries and +2.54 in hand-object interactions.

Explore all the technical details here: https://egovis.github.io/cvpr26/