EPIC-KITCHENS-100
100 hours of unscripted egocentric kitchen activities. The gold standard for action recognition and anticipation research.
Key Stats
| Metric | Value |
|---|---|
| Duration | 100 hours of unscripted kitchen activity |
| Environments | 45 kitchens across 4 cities |
| Action segments | 90K with verb-noun annotations |
| Size | ~750 GB |
| Modalities | RGB video, optical flow, object detections, action labels |
| License | CC-BY-NC-4.0 (non-commercial only) |
| Origin | University of Bristol |
What is EPIC-KITCHENS?
EPIC-KITCHENS-100 is the extended version of the EPIC-KITCHENS egocentric video dataset, capturing 100 hours of unscripted daily kitchen activities recorded by head-mounted cameras. Unlike staged datasets, participants simply went about their normal cooking and kitchen routines, producing naturalistic action sequences that reflect real-world manipulation patterns.
The dataset includes 90K fine-grained action segments annotated with verb-noun pairs (e.g., "cut tomato", "open drawer", "pour water"), making it the standard benchmark for egocentric action recognition, action anticipation, and cross-modal retrieval. Pre-extracted features (optical flow, object detections) are also available, reducing the compute barrier for researchers.
Relevance to robotics
EPIC-KITCHENS is widely used to train perception modules for kitchen manipulation robots. The action anticipation benchmark is directly relevant to predictive robot control -- if a robot can anticipate human actions, it can collaborate more effectively. The hand-object interaction patterns inform grasp planning and task sequencing in household robotics.
Access
Note: This dataset is CC-BY-NC-4.0 licensed. It is free for research but cannot be used commercially or redistributed for commercial purposes.
Related datasets
- Ego4D -- much larger egocentric dataset (3,670 hours)
- Assembly101 -- egocentric assembly with multi-view cameras
- RoboCasa -- simulated kitchen manipulation