EgoMAGIC: Egocentric Medical Video Dataset for AI Training

The EgoMAGIC dataset, detailed in arXiv:2604.22036, is an egocentric collection of medical activity videos gathered as part of DARPA's Perceptually-enabled Task Guidance (PTG) initiative. It features 3,355 videos covering 50 medical tasks, with a minimum of 50 labeled videos for each task. The goal of the PTG program was to create virtual assistants that work with augmented reality headsets to aid users in executing intricate tasks. This dataset presents an action detection challenge centered on eight medical tasks. Most recordings were captured using a head-mounted stereo camera with built-in audio. From this dataset, 40 YOLO models were developed, utilizing 1.95 million labels to identify 124 medical objects, establishing a solid foundation for medical AI advancements.