Dataset

Ego-Exo4D is a multi-modal multi-view video dataset and benchmark challenge centered on capturing egocentric and exocentric videos of skilled human activities. It supports multimodal machine perception research in daily life activities. The dataset was collected by 839 camera-wearing volunteers in 13 cities around the world, capturing 1,422 hours of video of skilled human activity. This dataset provides three natural language datasets such as expert comments, tutorial-style narratives provided by participants, and atomic action descriptions in one sentence, paired with videos. Ego-Exo4D also captures multiple viewpoints and multiple perception modalities, including multiple viewpoints, seven microphone arrays, two IMUs, a barometer, and a magnetometer. Datasets were recorded in strict compliance with privacy and ethics policies and with formal consent from participants. For more information, please visit the official website.