This dataset holds infrared recordings of contributors' hands and rooms, made with a Steam Frame headset. It's licensed under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0): non-commercial use only, with attribution. Contributors also grant Frametop's maintainer a separate non-exclusive license to their recordings for any purpose. By getting access you also agree: not to use the recordings to identify, recognise, track or profile anyone, or for biometric identification; not to try to find out who contributed or where a recording was made; not to publish images from the dataset that show identifiable details of a person or a home; and to delete recordings that a later version of the dataset withdraws. Contributors must be 18 or older, must not live in Illinois, Texas or Washington (USA), and must have the right to share what their recordings show. Read the full card before you agree.

Frametop Hands

Infrared recordings of people's hands from the tracking cameras of the Valve Steam Frame headset, made with Frametop's guided hand recorder. The dataset exists to train a small hand-tracking model that runs on the headset itself, for Frametop's hand cutouts over its virtual screens.

Every contribution comes from a person who recorded their own hands, reviewed every frame, deleted what they didn't want to share, and uploaded it themselves. The maintainer reviews each one for privacy before it's merged.

Frametop is an independent open-source project. It isn't affiliated with or endorsed by Valve or Hugging Face. Steam and Steam Frame are trademarks of Valve Corporation. The terms here were written in good faith without a lawyer; if something seems wrong, please open a discussion.

Datasheet

Motivation

Why it exists. Off-the-shelf hand models were trained on colour photos from normal lenses. The Steam Frame's tracking cameras are monochrome, infrared and fisheye. Those models track well in dim rooms and poorly in bright light on this headset. A model trained on the headset's own images should close that gap.
Who made it. Frametop's maintainer, DeeJanuz, with recordings from Frametop users who chose to take part.
Funding. None.

Composition

Instances. A session is one recording round by one contributor, in one lighting condition. A session has up to 11 takes, one per section of the guided script (below). A take holds frame sets at about 10 a second. A frame set is one image from each camera, taken at about the same moment.
Cameras. The headset's 4 tracking cameras, in 8-bit grey:

Camera Size Notes

slam_left, slam_right 1056 x 1024 lower, wide fisheye, infrared

upper_left, upper_right 640 x 480 upper, infrared

Some sessions may also hold other camera streams the recorder captured (each named in the set header). The cameras face outward: they don't see the wearer's face or eyes.
Also recorded, per take.
- prompts.jsonl: what the contributor was asked to do, and when (pose, hand, distance, object).
- poses.jsonl: the headset's position, about 250 times a second, and the controllers' when the contributor wore them (null otherwise).
- take.json: the take's status, the ranges the contributor deleted, and the camera clock's offset.
Per session.
- calibration.json: the cameras' intrinsics and extrinsics, with serial numbers and other unit identifiers removed.
- device.json: where the cameras sit on the head.
- manifest.json: the contributor's checklist answers (lighting, objects at hand, sleeves, rings, watch, whether controllers were worn), software versions, and a summary of the takes.
Labels. None from the contributors. The prompts say what the hands should be doing, and the controller sections give metric ground truth for depth. Keypoint labels are made later by an auto-labeller (a teacher model on every view, triangulation, a skeleton fit). They aren't part of this repository yet.
Size. About 8 GB per lighting round, compressed. This is an estimate from the recorder's export: 4 cameras at 10 sets a second come to about 28 MB a second before compression.
People in the data. Each contributor's hands, arms and clothing, and whatever was in view: their room, desk and the things on it. Contributors were asked to keep other people, mirrors, papers and screens out of view, and the maintainer checks for them, but the images can still show identifiable details of a person's home.
Identifiers. Each contributor is a random id (a UUID made on their headset). No names, emails, account names, headset serials or sound are recorded. Session times carry the contributor's time zone. The Hugging Face account that opened a pull request is visible on that pull request, next to the contributor id, but not in the data.

File layout

contributions/<contributor id>/<session YYYYMMDD-HHMMSS>/
  manifest.json
  calibration.json
  device.json
  SHA256SUMS
  takes/<NN>-<section>/
    sets.bin.zst      the frame sets: FHSET01 records, zstd-compressed
    prompts.jsonl
    poses.jsonl
    take.json

FORMAT.md gives the exact format of every file, and a short Python reader.

Collection process

The Frametop hand recorder guides the contributor through a script with a panel inside the headset. The contributor steps through it with the headset's button. A full round takes about 8-10 minutes, more with the optional sections:

Hand size: both hands at about 40 cm, flat, backs, then spread.
Pose sweeps: both hands, then each alone, moving near and far and all around while changing shape with the lit pictures: open, fist, point, pinch, OK, thumbs up, spread, claw, counting 1-5.
Gestures: pinch taps, pinch-drag, grabs, hands crossing and overlapping, hands near the face.
Desk work: typing and using the mouse.
Objects the contributor ticked: pen, phone, cup, keyboard, mouse, gamepad, small objects (optional).
Touch the dot: touching virtual dots at known 3D positions.
Controller depth with Frame controllers strapped on (optional).
Bridge: one controller on, the bare fingertip touching it (optional).
Bare push: pushing the open hands out and back, at several heights, without controllers.
No hands: about 10 s with the hands out of view.

Older sessions may have static-poses and wrist-rotations takes instead of pose-sweeps.

Rounds repeat in three lighting conditions: dim (lamp only), normal room light, and daylight near a window. A shorter core script is used after the first round.

The recorder runs on the headset and never uploads anything itself. The contributor reviews every take, can delete any stretch, take or session, exports what's left, and uploads it as a pull request with their own Hugging Face account.

Consent and privacy

Contributors read and accept the consent text in the recorder before their first session, and confirm they're 18 or older and don't live in Illinois, Texas or Washington (USA), whose biometric privacy laws need more than this project can provide yet. The text is in Frametop's repository: hands/rec/CONSENT.md.
They're asked to face away from other people, mirrors, papers and screens, and to make sure no one else's face or hands are in view.
Before a pull request is merged, the maintainer checks its files and looks through contact sheets of every take, from every camera, for faces, other people, papers, readable screens, and anything else private. A pull request that fails is closed and its uploaded files are deleted.
Free-text fields (notes, own objects) are read by the maintainer before merging.
The recordings are used to train and test hand-tracking models. They aren't used to identify people, and the access terms forbid anyone from trying.
Contributors may have further rights over their data depending on where they live (for example under the GDPR); requests go to the maintainer (see Maintenance).

License and contributor grant

The dataset: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). You may share and adapt it for non-commercial purposes, with attribution. Credit contributors by their contributor id, and cite the dataset (below).
The contributor grant: each contributor also gives Frametop's maintainer (today DeeJanuz, and whoever takes over maintaining Frametop) a non-exclusive, worldwide, royalty-free license to use their contribution for any purpose, commercial included: copying, changing, and training, publishing and selling models made from it. It doesn't give anyone else those rights.
Privacy isn't licensed. CC BY-NC 4.0 covers copyright and database rights only; it doesn't license anyone's privacy or publicity rights. The access terms above apply on top of it.
Models trained on this data aren't covered by this card. Whether model weights are a derivative of their training data is legally unsettled; Frametop's own models come with their own license.
See LICENSE.

How to contribute

See CONTRIBUTING.md. In short: install Frametop on a Steam Frame, open the hand recorder, record, review, export, then log in and upload on the recorder's Upload page. The upload is a pull request from your Hugging Face account.

Withdrawal

A contributor can withdraw at any time, without giving a reason. Before a pull request is merged, close it. After, open a discussion on this dataset, or an issue in Frametop's repository, with your contributor id (shown in the recorder) and which sessions, or "all". Post from the Hugging Face account that uploaded them if you can, or say how the maintainer can check it's you. Don't post anything else that identifies you.

Then, usually within 30 days, the maintainer:

deletes those recordings from the dataset in one commit;
squashes the repository's history, so no older commit holds them, and permanently deletes their stored files, including those in pull-request branches;
leaves them out of anything trained after that.

This rewrites the repository's history: commit ids change, and anyone who pinned a revision or cloned the repository needs to fetch it again. If you use this dataset, please update your copy when a new version removes recordings, and delete the withdrawn files you hold. Copies made before a withdrawal, and models trained before it, can't be recalled.

Maintenance

Maintainer: DeeJanuz, through this dataset's discussions or Frametop's issues.
Updates: contributions are merged as they're reviewed. Withdrawals are handled as they arrive.
Errors: report a problem with a recording (something private in view, a broken file) on the discussions page, quoting the contributor id, session and take, without describing what you saw in detail.

Known limitations

Few people. Early versions hold recordings from a handful of contributors, all Frametop users. They aren't a balanced sample of hand sizes, skin tones, ages, or accessories.
One headset model. Only the Steam Frame's cameras. Each unit's calibration differs a little.
Infrared only. Monochrome infrared images; ambient light changes how they look a lot.
10 frames a second. Fast motion is blurred and gaps between frame sets are about 100 ms.
Scripted motion. Poses follow the prompts. Natural, unprompted hand use is under-represented.
No labels yet. Prompts say what the hands should be doing; contributors don't always do exactly that.
Private rooms. Contributors' homes are in view. Don't publish images from this dataset that show identifiable details of a room, and don't try to find out where it is.
Withdrawals. Recordings can disappear in later versions. Pin a revision for reproducibility, and report results on the version you used.

Citation

@misc{frametop_hands,
  title        = {Frametop Hands: infrared hand recordings from the Steam Frame},
  author       = {DeeJanuz and the Frametop Hands contributors},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DeeJanuz/frametop-hands}},
  note         = {CC BY-NC 4.0}
}

Downloads last month: 8

Total file size:

6.63 GB

Camera	Size	Notes
`slam_left`, `slam_right`	1056 x 1024	lower, wide fisheye, infrared
`upper_left`, `upper_right`	640 x 480	upper, infrared