humyn-labs/APAC-Egocentric-Monocular-Labeled

APAC Egocentric Monocular (Labeled)

Twelve labeled egocentric work sequences from industrial, hospitality, logistics and retail settings, released as hand-tracking and head-tracking renders with 252 densely captioned 3-second action segments.

Preview: polishing a car in an automotive garage — hand-tracking render, downscaled to 720p.

---

At a glance

Samples	12
Total duration	12 min 09 s (~61 s each)
Resolution	1920×1080 @ 30 fps
Action segments	252 (~21 per clip, 3 s each)
Unique verbs	87
Audio	none
Environments	Industrial, Hospitality, Logistics, Retail (3 scenes each)

Scenes

Environment	Scenes
Industrial	Assembly Line · Automotive Garage · Construction
Hospitality	Pantry · Housekeeping · Kitchen
Logistics	Storage Hub · Electronics Factory · Laundromat
Retail	Clothes Store · Accessories Store · Home Decor Shop

---

Important: this release contains tracking renders, not raw camera footage

Each sample ships as two rendered views — a hand-tracking video and a head-tracking video — with tracking overlaid on the scene. The unmodified egocentric camera feed is not part of this release. If your method needs pristine RGB input, this dataset will not provide it; see the stereo release, which does include rectified raw video.

---

Repository layout

data/train-*.parquet   # 720p previews + all metadata + action segments (powers the viewer)
hand_tracking/*.mp4    # full-resolution hand-tracking renders (1920×1080)
head_tracking/*.mp4    # full-resolution head-tracking renders
annotations/*.jsonl    # action labels, one segment per line
preview/*.mp4          # 720p proxies used by the viewer and this card
metadata.csv           # flat table

The parquet embeds the 720p previews, not the originals, so the viewer stays fast and every clip plays in the browser. Full-resolution renders are in hand_tracking/ and head_tracking/.

Columns

Column	Description
video	720p hand-tracking preview — plays in the viewer
head_tracking_preview_video	720p head-tracking preview
sample_id	UUID shared by every file for this sample
environment, scene	Industrial/Hospitality/Logistics/Retail, and the specific setting
task_description	What the operator is doing
duration_seconds, width, height, fps	Probed from the media
hand_tracking_path, head_tracking_path	Full-resolution files in this repo
action_labels_path	The .jsonl for this sample
num_segments	Labeled segments in this clip
verbs, objects	Distinct verbs and object names in the clip
segments	Every labeled segment, inline — see below
hand_tracking_sha256, hand_tracking_bytes	Integrity and size of the original

Action label format

One JSON object per 3-second segment:

{"video_id": "0114fd12-c0fb-4115-9cf5-4285e6516f7d",
 "segment_id": 0,
 "segment_start_time": "00:00.0", "segment_end_time": "00:03.0",
 "verb": "polishing",
 "object_name": "car hood", "object_color": "black", "object_shape": "curved",
 "object_context": "parked in a garage",
 "object_description": "black curved car hood parked in a garage",
 "caption": "The operator is polishing a black car hood with an electric buffer. The left hand guides the tool's head while the right hand controls the handle, moving the spinning pad across the surface."}

Captions are two-sentence descriptions that name the acting hand and the tool, which makes them usable for bimanual manipulation grounding rather than just clip-level classification.

---

Usage

from datasets import load_dataset

ds = load_dataset("humyn-labs/APAC-Egocentric-Monocular-Labeled", split="train")
r = ds[0]
print(r["environment"], r["scene"], r["task_description"])
print(r["segments"][0]["caption"])

Every segment across the corpus:

segs = [s for r in ds for s in r["segments"]]
print(len(segs), "segments")

Full-resolution renders:

from huggingface_hub import snapshot_download
snapshot_download("humyn-labs/APAC-Egocentric-Monocular-Labeled", repo_type="dataset",
                  allow_patterns=["hand_tracking/*", "head_tracking/*", "annotations/*"])

Intended uses

Fine-grained action recognition · dense video captioning · verb–object grounding · hand–object interaction · industrial and retail workflow understanding · vision-language pretraining for embodied agents.

Limitations

• Small. 12 clips, ~12 minutes total — an evaluation and prototyping pack, not pretraining scale.
• Tracking renders only, as described above.
• One clip per scene, so there is no within-scene variation to train on and no held-out split is provided.
• Uniform length (~61 s) means no long-horizon sequences.
• Segment boundaries are on a fixed 3-second grid, not action-aligned, so a labeled segment can straddle two actions.
• No audio, no depth, no camera pose.

Provenance

Curated from the HumynLabs egocentric sample collection. All technical fields were probed from the media rather than copied from the source sheet. Every listed file resolved and downloaded; no samples were dropped.

License

CC BY 4.0. Recorded in working environments with participant consent for research use. Bystanders and workplace interiors may appear incidentally — please do not attempt to re-identify individuals.

Citation

@misc{humynlabs2026apacmonocular,
  title  = {APAC Egocentric Monocular (Labeled)},
  author = {HumynLabs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/humyn-labs/APAC-Egocentric-Monocular-Labeled}
}