EgoViz-120

120 hours of real-world egocentric data, recorded on the job.

Data bucket · Humaid · humaid.co · info@humaid.co

Trained operators wearing a head-mounted stereo rig and two wrist cameras, doing their actual work: mopping corridors, prepping food, cleaning restrooms, running laundry. Nothing staged, nothing re-shot. Every clip is one continuous activity with all sensors on a shared clock, in a single self-contained MCAP.

_{One clip's preview render — ego stereo, both wrists, depth, head odometry, three IMUs, the projected hand skeletons, and the action label as it changes.}

Where the data lives. The dataset is 7.25 TB, so it is served from a Hugging Face Storage Bucket: hf://buckets/humaidtech/EgoViz-120. This repository is the card and the entry point — see Getting the data.

At a glance

Duration	120.4 hours
Clips	5,392
Categories	12
Size	7.25 TB — 6.81 TB MCAP, 445 GB MP4
Clip length	15 s min · 44 s median · 80 s mean · 35 min max
License	CC-BY-4.0

A clip

data/<category>/<chunk_id>/
    <chunk_id>.mcap    every sensor stream + frame-level action labels
    <chunk_id>.mp4     1920×1080 dashboard preview
    <chunk_id>.json    clip metadata

The MCAP — 23 channels on one clock

ROS 2 profile, zstd-compressed.

Stream	Topic	Rate	Type
Ego stereo	/ego/zed_head/{left,right}/image/compressed	30 Hz	foxglove.CompressedVideo (H.264)
Ego depth	/ego/zed_head/depth/image/compressed	30 Hz	sensor_msgs/msg/CompressedImage (16-bit PNG)
Wrist cameras	/wrist_{left,right}/image/compressed	30 Hz	foxglove.CompressedVideo (H.264)
IMUs ×3	/ego/zed_head/imu/data, /wrist_{left,right}/imu/data	~200 Hz	sensor_msgs/msg/Imu
Hand keypoints	/hands/{left,right}/keypoints	30 Hz	visualization_msgs/msg/MarkerArray
Hand pose	/hands/{left,right}/odom	30 Hz	nav_msgs/msg/Odometry
Hand presence	/hands/{left,right}/detected	30 Hz	std_msgs/msg/Bool
Head odometry	/ego/zed_head/odom	30 Hz	nav_msgs/msg/Odometry
Actions	/action	30 Hz	tad_msgs/msg/ActionPhase
Transforms	/tf, /tf_static	30 Hz	tf2_msgs/msg/TFMessage
Calibration	*/camera_info ×5	30 Hz	sensor_msgs/msg/CameraInfo

All three cameras are 1920×1200. Depth is stored as lossless 16-bit grayscale PNG.

The preview MP4

The video at the top of this page is one. A 1920×1080 render of everything at once — ego left and right, both wrists, colorized depth, the 3D odometry trail, three IMU traces, projected hand skeletons, and the current action with a progress bar. Enough to browse the dataset without decoding a single MCAP.

The JSON

{
  "chunk_id": "77e637de67a572c9a635cdf2273cdf59",
  "category": "preparing_cleaning_supplies",
  "title": "Washing cleaning cloth",
  "description": "The worker washes a cleaning cloth in the sink.",
  "duration_s": 15.0,
  "group_id": "g001",
  "sequence_id": "s162",
  "index": 9,
  "start_ts_ms": 545000,
  "end_ts_ms": 560000
}

group_id is the capture session. Clips sharing one are the same worker, site and lighting, so split on group_id to avoid leakage between train and test. sequence_id is the continuous recording within that session and index is the clip's position in it, so (sequence_id, index) reconstructs the original order; start_ts_ms and end_ts_ms are offsets within that recording.

Categories

Category	Clips	Hours
mopping_floor	1,595	56.8
wiping_surfaces	777	14.7
preparing_food	541	13.4
preparing_cleaning_supplies	790	8.5
cleaning_restroom	332	5.7
cleaning_windows	172	4.3
organizing_items	279	3.8
sweeping_floor	159	3.4
cooking_food	193	2.8
managing_laundry	210	2.5
managing_trash	208	2.2
cleaning_kitchen	136	2.1

Getting the data

pip install -U huggingface_hub
hf auth login

B=hf://buckets/humaidtech/EgoViz-120

Start with the index. One row per clip — category, title, description, duration, skill and object lists, and bucket paths. Under 1 MB, and it answers most questions before you download any video.

hf buckets cp $B/index.parquet index.parquet

One clip, all three files:

CID=77e637de67a572c9a635cdf2273cdf59
hf sync $B/data/preparing_cleaning_supplies/$CID ./$CID

One category:

hf sync $B/data/cleaning_kitchen ./cleaning_kitchen

Previews and metadata only, skipping the MCAPs:

hf sync $B/data ./egoviz --include "*.mp4" --include "*.json"

The bucket also speaks S3 at https://s3.hf.co/humaidtech, so rclone, s5cmd, boto3 and the AWS CLI work against it — see S3 compatibility.

How it was built

1. Capture — head-mounted ZED stereo camera, two wrist cameras and three IMUs, recorded on-body during ordinary shifts.
2. Segmentation — continuous recordings split into single-activity clips, each with a category, title and description.
3. Hand pose — 3D keypoints and per-hand odometry recovered for both hands at 30 Hz.
4. Action labeling — the ego-left stream was decomposed into actions and atomic phases, written back into the MCAP as /action.

Privacy

Camera wearers are trained operators who consented to recording and to public release. All footage — the MCAP video streams and the preview MP4s alike — went through an automated face-blurring pass before publication, covering operators and anyone appearing in the background.

For privacy enquiries, write to info@humaid.co.

Limitations

• Category imbalance. mopping_floor is 47% of all hours. Sample accordingly.
• Labels are model-generated and not human-verified. The skill and object vocabularies are open rather than a fixed taxonomy, so expect near-duplicates (mop / mopping).
• Single geography. Lighting, layout and tooling conventions reflect one region.

Citation

@misc{egoviz120_2026,
  title        = {EgoViz-120: 120 Hours of Synchronized Real-World Egocentric Data},
  author       = {Humaid Tech Inc.},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/humaidtech/EgoViz-120}}
}

Contact

Open a discussion on this repository, or write to info@humaid.co.

We build datasets to spec — environments, tasks, sensor configuration, annotation schema and volume. If real-world data is your bottleneck, get in touch.