fira-aslan/epfl-smart-kitchen-hands-pose

EPFL-Smart-Kitchen hand videos, 3D pose and camera viewpoints (train / test)

Hand-crop video segments, full 3D body and hand pose, and two static camera viewpoints from the EPFL-Smart-Kitchen-30 dataset, packaged as zips per split.

Files

file	size	contents
train.zip	28.2 GB	29,137 hand-crop segments + pose for 26 sessions
test.zip	17.9 GB	17,484 hand-crop segments + pose for 16 sessions
train_views.zip	2.0 GB	3,844 viewpoint clips, 26 sessions
test_views.zip	1.2 GB	2,408 viewpoint clips, 16 sessions

train.zip / test.zip:

videos/00000000.mp4 ...                       hand-crop segments, one per annotation row
pose_data/<participant>/<session>/            pose3d_mano_wconf.csv (hands),
                                              pose3d_smpl_wconf.csv (body),
                                              pose_3d_wconf/, pose_3d_wconf_all/
annotations/<split>_data_dedup.csv
annotations/split_<split>_new.txt

train_views.zip / test_views.zip:

views/<participant>/<session>/Aoutput1/<start_sec>.mp4
views/<participant>/<session>/Boutput3/<start_sec>.mp4
views_manifest.csv                            one row per clip
annotations/<split>_data_dedup.csv            same as above plus a view_chunks column
annotations/split_<split>_new.txt

Viewpoint clips are contiguous non-overlapping 30-second chunks of the session recording, 512x288 at 30 fps, named by their start second (0.mp4, 30.mp4, 60.mp4, ...). Aoutput1 and Boutput3 are two static cameras; both cover the identical set of chunks in every session.

Joining segments to videos

• Hand crop: the video column of the annotations CSV gives the filename directly.
• Viewpoints: the view_chunks column lists the semicolon-separated start seconds of the chunks covering that segment, e.g. 120;150 when a segment crosses a 30 s boundary. The same chunk names exist under both cameras.

Do not join on segment_id — it is a 0..N-1 renumbering of the deduplicated rows and does not track the original clip indices.

Deduplication

The upstream train_data_new.csv and test_data_new.csv contain exact duplicate rows (58,324 -> 29,137 and 35,032 -> 17,484). Duplicates were dropped keeping the first occurrence. The duplicated rows were verified to point at byte-identical video files, so no distinct footage is lost by dropping them.

Splits are session-level and disjoint by session. The validation split is not included.

Source and citation

Derived from the EPFL-Smart-Kitchen-30 dataset, released under CC BY 4.0.

Bonnetto et al. (2025). EPFL-Smart-Kitchen-30: Densely annotated cooking dataset with 3D kinematics to challenge video and language models.

This derivative is redistributed under the same CC BY 4.0 licence.