Egocentric Vision

Egocentric vision studies what a person sees and does through a wearable, first-person camera, including recognizing activities, tracking hands and body motion, locating objects, and predicting what happens next. Methods combine video models with temporal reasoning, 3D geometry, stereo or event cameras, and signals such as gaze, inertial measurements, and muscle activity; datasets and low-cost capture systems provide real-world training and evaluation. Recent work also uses generative models and world models to simulate interactions and transfer human demonstrations to robots. A central challenge is reliability under occlusion, rapid camera motion, poor lighting, and changing environments, while preserving physical and scene state over long tasks.