EgoMAGIC- An Egocentric Video Field Medicine Dataset for Training Perception Algorithms
3,355 egocentric videos of 50 medical tasks with 1.95 million labels now public
A team of researchers from DARPA's Perceptually-enabled Task Guidance (PTG) program has released EgoMAGIC (Medical Assistance, Guidance, Instruction, and Correction), a large-scale egocentric video dataset designed to train perception algorithms for field medicine. The dataset comprises 3,355 videos capturing 50 distinct medical tasks, with a minimum of 50 labeled examples per task. Recorded primarily using head-mounted stereo cameras with integrated audio, the dataset provides a first-person perspective ideal for training AI assistants that operate in real-world, hands-on medical scenarios.
To accelerate development, the team trained 40 YOLO models using 1.95 million labels covering 124 medical objects, offering a robust starting point for medical AI applications. Baseline results on action detection for eight selected tasks achieved an average mAP of 0.526. Beyond action detection, EgoMAGIC supports action recognition, object identification, error detection, and other computer vision challenges. The dataset is publicly available via Zenodo (DOI: 10.5281/zenodo.19239154).
- 3,355 egocentric videos covering 50 medical tasks with at least 50 labeled videos per task
- 1.95 million labels across 124 medical objects, with 40 pre-trained YOLO models
- Baseline action detection achieves average mAP 0.526 on eight selected medical tasks
Why It Matters
Open-source medical video dataset enables AR-based AI assistants for real-time field medicine training and error detection.