EPIC-KITCHENS

A classic first-person video dataset of unscripted everyday cooking and kitchen tasks, filmed with head-mounted cameras.

EPIC-KITCHENS is a first-person video dataset led by the University of Bristol in the UK. The first version, EPIC-KITCHENS-55, was released in 2018 with 55 hours from 32 participants. It was expanded in 2020 into EPIC-KITCHENS-100: participants wore head-mounted cameras while cooking and washing dishes in their own kitchens, unscripted, across 45 kitchens, totaling 100 hours, 20 million frames, and about 90,000 action segments. Annotations come from participants narrating afterward what they were doing, which is then organized into “verb + noun” action labels. Companion benchmarks cover action recognition, detection, anticipation, cross-modal retrieval, and unsupervised domain adaptation. Later additions layered on hand-object segmentation (VISOR), sound events (EPIC-Sounds), and 3D camera information (EPIC-Fields). In robotics, it's commonly used to study hand-object interaction, affordances, and pretraining visual representations.

ExampleOne action segment is labeled “take plate”: the video shows the participant lifting a plate off a shelf, and the label is built from the verb take and the noun plate.

Also called

EPIC-KITCHENS-100, EPIC-KITCHENS-55

Related

Egocentric Video · Ego4D · Human Video Data · Hand-Object Interaction · Affordance Detection · Data Annotation

Sources

Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100 (arXiv)
The EPIC-KITCHENS Dataset: Collection, Challenges and Baselines (arXiv)
EPIC-KITCHENS 官网 (Chinese)

As of

2025-06