dataset link on Hugging dataset
https://huggingface.co/datasets/lightly-ai/epic-kitchens-100-clips
Arxiv link
https://arxiv.org/pdf/2006.13256
Description of the dataset
Part of the MOEB tracker (#4842) — picking up the "video clustering" task-type gap assigned to me and @wissam-sib.
EPIC-KITCHENS-100 is a large-scale egocentric (first-person) video dataset of unscripted daily kitchen activities, annotated with verb/noun/action narration labels via a "Pause-and-Talk" interface (Damen et al., IJCV 2021 / TPAMI 2021). The official raw release is very large (740GB–1.1TB) and distributed via Data.Bris/Academic Torrents — impractical to mirror in full for a benchmark task.
Using the lightly-ai/epic-kitchens-100-clips HF mirror instead: 37,455 pre-trimmed action-level clips (24GB total, downscaled/compressed), each corresponding to exactly one narration annotation ({participant_id}/{narration_id}.mp4), with the official EPIC_100_train/validation annotation CSVs bundled in the same repo. CC-BY-NC-4.0, not gated, publicly downloadable (verified directly).
Plan: sample a benchmark-sized subset spanning a good spread of verb/noun/action categories, use the action-class labels as cluster ground truth (same pattern as the existing UrbanSound8K/BirdCLEF audio clustering tasks — labels become the clusters, no relevance-rule invention needed). Will include the construction script and exact sample counts in the PR.
dataset link on Hugging dataset
https://huggingface.co/datasets/lightly-ai/epic-kitchens-100-clips
Arxiv link
https://arxiv.org/pdf/2006.13256
Description of the dataset
Part of the MOEB tracker (#4842) — picking up the "video clustering" task-type gap assigned to me and @wissam-sib.
EPIC-KITCHENS-100 is a large-scale egocentric (first-person) video dataset of unscripted daily kitchen activities, annotated with verb/noun/action narration labels via a "Pause-and-Talk" interface (Damen et al., IJCV 2021 / TPAMI 2021). The official raw release is very large (740GB–1.1TB) and distributed via Data.Bris/Academic Torrents — impractical to mirror in full for a benchmark task.
Using the
lightly-ai/epic-kitchens-100-clipsHF mirror instead: 37,455 pre-trimmed action-level clips (24GB total, downscaled/compressed), each corresponding to exactly one narration annotation ({participant_id}/{narration_id}.mp4), with the official EPIC_100_train/validation annotation CSVs bundled in the same repo. CC-BY-NC-4.0, not gated, publicly downloadable (verified directly).Plan: sample a benchmark-sized subset spanning a good spread of verb/noun/action categories, use the action-class labels as cluster ground truth (same pattern as the existing UrbanSound8K/BirdCLEF audio clustering tasks — labels become the clusters, no relevance-rule invention needed). Will include the construction script and exact sample counts in the PR.