EgoLife
EgoLife is a large-scale egocentric video dataset designed to train models to detect and understand human activities in a daily context. It offers annotated first-person videos for tasks such as action recognition, scene analysis or even temporal segmentation.
Description
EgoLife is an egocentric video dataset of more than 30,000 clips captured in real everyday contexts (home, kitchen, transport, etc.). The videos are accompanied by precise annotations on the actions, objects present, activity transitions, etc. The corpus is designed for modeling human behaviors and training computer vision systems focused on human activity.
What is this dataset for?
- Train video activity recognition models (action recognition)
- Analyzing daily life sequences for research in cognition or robotics
- Evaluating multimodal models in real life contexts
Can it be enriched or improved?
Yes, it is possible to add fine annotations (duration of the action, social interactions), or to cross-reference the videos with audio or physiological data. The open format of the dataset also makes it possible to create customized subsets according to the tasks (eg. object detection, temporal segmentation, etc.).
🔎 In summary
🧠 Recommended for
- Computer vision researchers
- Robotics projects
- Behavioral analysis
🔧 Compatible tools
- PyTorchVideo
- MMAction2
- TensorFlow
- Detectron2
💡 Tip
Use tools like Chronos or DVC to better manage the storage and preprocessing of large video clips.
Frequently Asked Questions
Are the videos annotated manually?
Yes, video clips are accompanied by human annotations about the actions, objects, and types of scenes observed.
Can this dataset be used for object detection in a first-person view?
Yes, although the main focus is on human activity, the videos also include annotated objects that can be used for detection.
Do you need specific resources to use videos?
It is recommended to have a GPU and sufficient storage space, as videos can be cumbersome to process.
