humyn-labs/APAC-Egocentric-Stereo-Labeled

APAC Egocentric Stereo (Labeled)

Twelve labeled egocentric work sequences captured with a stereo rig, each shipping rectified side-by-side raw video, a depth map, a hand/head-tracking render, and 248 densely captioned 3-second action segments.

Unlike the monocular release, this set includes unmodified rectified camera footage, so it supports stereo matching and depth research directly.

Preview: removing a tail lamp in an automotive garage — rectified side-by-side stereo pair (left and right eye shown together), downscaled to 720p.

---

At a glance

Samples	12
Total duration	12 min 01 s (~60 s each)
Rectified SBS video	3840×1080 @ 30 fps (1920×1080 per eye)
Depth maps	1920×1080
Action segments	248 (~21 per clip, 3 s each)
Unique verbs	78
Audio	none
Modalities per sample	Stereo RGB · depth · hand/head tracking · action labels

Scenes

Environment	Scenes
Industrial	Automotive Garage · Construction Site · Electronics Factory
Hospitality	Restaurant · Kitchen · Bar
Logistics	Shipment Hub · Electronics Factory · Laundromat
Retail	Shopping Mart · Clothes Store · Book Store

---

Repository layout

data/train-*.parquet   # 720p previews + all metadata + action segments (powers the viewer)
videos/*_sbs_rectified.mp4   # full-resolution rectified stereo pairs (3840×1080)
depth/*_depth_map.mp4        # depth maps (1920×1080)
tracking/*_hand_tracking.mp4 # hand/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 720p previews, not the 8 GB of originals, so the viewer stays fast and every clip plays in the browser. Full-resolution media lives in videos/, depth/ and tracking/.

Splitting the stereo pair

The SBS frame is the left and right eye concatenated horizontally. To recover each eye:

import av, numpy as np

frame = next(av.open("videos/<id>_sbs_rectified.mp4").decode(video=0)).to_ndarray(format="rgb24")
left, right = frame[:, :1920], frame[:, 1920:]   # 1920×1080 each

or with ffmpeg:

ffmpeg -i in_sbs_rectified.mp4 -filter_complex "[0]crop=1920:1080:0:0[l];[0]crop=1920:1080:1920:0[r]" -map "[l]" left.mp4 -map "[r]" right.mp4

Columns

Column	Description
video	720p SBS preview — plays in the viewer
depth_preview_video, tracking_preview_video	720p depth and tracking previews
sample_id	UUID shared by every file for this sample
environment, scene, task_description	Setting and what the operator is doing
sbs_width, sbs_height, per_eye_width	3840, 1080, and 1920 respectively
depth_width, depth_height, fps, duration_seconds	Probed from the media
sbs_video_path, depth_map_path, tracking_path, action_labels_path	Full-resolution files in this repo
num_segments, verbs, objects	Segment count and distinct verbs/objects
segments	Every labeled segment, inline
sbs_sha256, sbs_bytes	Integrity and size of the original

Action label format

One JSON object per 3-second segment, identical in schema to the monocular release:

{"video_id": "27c96099-7718-437b-8e91-ef6d6e8591aa",
 "segment_id": 0,
 "segment_start_time": "00:00.0", "segment_end_time": "00:03.0",
 "verb": "unscrewing",
 "object_name": "screw", "object_color": "metallic", "object_shape": "cylindrical",
 "object_context": "mounted on the car's taillight housing",
 "object_description": "metallic cylindrical screw mounted on the car's taillight housing",
 "caption": "The person uses a screwdriver to loosen a screw on the car's red taillight. The left hand guides the tip while the right hand rotates the handle."}

---

Usage

from datasets import load_dataset

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

Full-resolution stereo, depth and tracking:

from huggingface_hub import snapshot_download
snapshot_download("humyn-labs/APAC-Egocentric-Stereo-Labeled", repo_type="dataset",
                  allow_patterns=["videos/*", "depth/*", "tracking/*", "annotations/*"])

Intended uses

Stereo depth estimation · self-supervised depth from egocentric video · hand–object interaction in 3D · fine-grained action recognition · dense video captioning · industrial workflow understanding · embodied vision-language pretraining.

Limitations

• Small. 12 clips, ~12 minutes — an evaluation and prototyping pack, not pretraining scale.
• No camera calibration shipped. The footage is rectified, but intrinsics, baseline and extrinsics are not included, so metric reconstruction is not possible from this release alone.
• Depth maps are provided as video, so they carry 8-bit compression artifacts and are not a substitute for raw sensor depth.
• One clip per scene, no within-scene variation and no held-out split.
• Segment boundaries sit on a fixed 3-second grid, not action-aligned.
• No audio, no IMU, 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{humynlabs2026apacstereo,
  title  = {APAC Egocentric Stereo (Labeled)},
  author = {HumynLabs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/humyn-labs/APAC-Egocentric-Stereo-Labeled}
}