Access is limited to hand-tracking evaluation and research. The footage may contain incidental personal information. Do not redistribute or publish it, do not attempt to identify anyone or any location, protect it appropriately, and delete it when access is revoked or MicroAGI requests deletion. By requesting access you accept DATA-USE-AGREEMENT for yourself and your organisation.

hand-tracking-evals

A shared benchmark for egocentric hand tracking: the sequences, the data contract, the prediction format, and the scorer — so two teams can compare hand trackers and have the numbers mean the same thing.

Run your tracker however you like, on whatever hardware you like, and submit a file. Scoring is deliberately separate from inference: nothing in handeval imports torch, loads a model or needs a GPU, so whoever ran the tracker, the scoring is identical code. Every participant, including MicroAGI, submits the same public prediction format. There is no private scoring path.

Gated repo. Access is granted per person after accepting the data-use agreement. The code is Apache-2.0; the footage is not — see Terms before doing anything with it.

Quickstart

pip install huggingface_hub
hf auth login
hf download MicroAGI-Labs/hand-tracking-evals --repo-type dataset --local-dir hand-tracking-evals

cd hand-tracking-evals && pip install -e .

# see the whole flow on synthetic data, no licence needed
python examples/make_demo_suite.py --out /tmp/demo --noise 0.01
python -m handeval run --manifest /tmp/demo/suite --predictions /tmp/demo/predictions \
                       --all --paper --out /tmp/demo/results

The demo predictions are ground truth plus a known error, so the result is checkable by hand: --noise 0.01 should read as roughly a 16 mm clean MPJPE. If it does not, the harness is lying to you.

Then on the real suite:

python -m handeval list     --suite phone_edge_cases --all
# run your tracker, write predictions/<sequence>.json   (docs/PREDICTIONS.md)
python -m handeval validate --suite phone_edge_cases --all --predictions predictions/
python -m handeval run      --suite phone_edge_cases --all --predictions predictions/ \
                            --out runs/mine

Run validate the first time you integrate. It checks that your frame indices line up with the sequences — the one mistake that produces perfectly valid files and quietly wrong numbers.

What is here

path	what
`handeval/`	the scorer. numpy + scipy, CPU only
`docs/PREDICTIONS.md`	the format — read this first. Coordinate frame, joint order, states
`docs/METRICS.md`	what every number means, and how to read it honestly
`docs/DATASETS.md`	the suites, and how to obtain each one
`docs/THIRD_PARTY_BUILD.md`	exact third-party IDs, input receipts, commands and validation
`docs/GT_MCAP.md`	the data contract, if you add a suite
`converters/`	portable third-party dataset → GT MCAP tools
`suites/phone_edge_cases/`	the one dataset we can ship — 51 clips (below)

The suites

phone_edge_cases is included — 51 nominal six-second clips of egocentric hand tracking going wrong, drawn from 21 production sessions because a human watched them fail. Each is a nominal six-second window at 1920×1080, ~29.98 Hz (179–180 frames; 5.935–5.972 s first-to-last timestamp span).

It has no ground truth and never will. What it measures is a method's self-reported stability, not accuracy: declared detections, fragmentation, and pose motion. Its purpose is --compare against a previous run of the same tracker under the same state semantics. A single number means nothing, and these metrics must never be used to rank different companies or model families.

Each clip carries one or more categories, in the manifest's cases, so a comparison can be sliced with --case:

category	clips	what the reviewer saw
`detection`	24	a hand that is present is not emitted, or one is emitted where there is none
`interaction_occlusion`	23	interacting hands, or a hand hidden by an object, the body or the task
`pose_fit`	11	the hand is found and placed, but the articulated pose is inaccurate
`interpolation`	10	the filled or held pose is wrong — flips, sweeps, bad bridging
`jitter`	8	temporally unstable output: finger wiggle, shake
`bystander`	5	another person's hands are in frame
`handedness`	4	left/right mislabelled, or the track swaps between hands

27 clips carry more than one. The labels record what a reviewer saw one tracker do wrong at that moment — a map of where to look, not ground truth, and not a claim about your tracker. Nothing in the scoring reads them.

A file is exactly its scored window, with no context padding. A tracker with warm-up state therefore starts cold on every clip. Everyone scoring this suite carries that equally, so run-to-run comparisons are fair; numbers from this suite are not comparable to the same tracker run over full sessions. The colour stream is re-encoded (libx265 CRF 16, IDR at frame 0) because the source keyframe interval is 250 frames and an exact-window copy would not decode standalone; every original frame timestamp is preserved.

The other seven suites include metadata, not data. HOT3D, HOI4D, ARCTIC, TACO, EgoDex, H2O and OakInk2 are third-party datasets whose licences do not permit redistribution — not the raw data, and not converted copies, which are derivative works. Exact manifests, upstream sequence IDs, split assignments and metadata hashes ship instead: accept each dataset's own licence, download it, run the documented conversion, validate the outputs, and point HANDEVAL_SUITE_ROOT at the result. Everyone then scores identical sequences with identical code — which is what makes a comparison mean anything — without anyone redistributing anything they may not. See docs/DATASETS.md for the per-dataset recipe.

Terms

Code (handeval/, converters/, docs/, synthetic examples) — Apache-2.0, see LICENSE-CODE.

Footage (suites/phone_edge_cases/clips/) — not Apache-2.0, and not covered by it. These are recordings of real people in real homes and workplaces, captured on a head-mounted camera during ordinary work.

Access is granted to named recipients for evaluation and research only under DATA-USE-AGREEMENT. In particular: do not train on or redistribute the footage, do not publish frames, do not attempt to identify a person or location, protect access, and delete it when access is revoked or MicroAGI requests deletion. Although the release is intended not to contain direct identifiers, incidental personal information may remain; report concerns to data@micro-agi.com rather than investigating them.

Data-use agreement

The following terms apply to the video and recorded data under suites/phone_edge_cases/. They do not apply to the software, documentation, or synthetic examples, which use Apache-2.0. By requesting, accessing, downloading, or using the evaluation data, you and the organisation on whose behalf you act agree:

Use the data only to evaluate, validate, compare, or research hand-tracking systems. Do not train or fine-tune a model on this held-out evaluation data.
Do not publish, sublicense, sell, transfer, share, or otherwise redistribute the data or extracted frames. Access is individual; collaborators must request their own access.
Employees, contractors, cloud providers, and subprocessors may process it only under confidentiality and data-protection obligations at least as protective as these terms. You remain responsible for their compliance.
You may share aggregate metrics, reports, and predictions that neither reproduce nor permit reconstruction of the data. Do not share images, video, audio, thumbnails, reversible embeddings, or other reconstructive derivatives.
The release is intended not to contain direct identifiers, but incidental personal information may remain. Do not identify, contact, profile, or infer sensitive attributes about a person, household, workplace, or location. Do not publish frames or recordings.
Protect the data with reasonable safeguards. Promptly report suspected unauthorised access, disclosure, or loss to data@micro-agi.com and reasonably cooperate in remediation.
Retain it only while approved access and a permitted purpose continue. Delete every copy when access is revoked, MicroAGI requests deletion, or that purpose ends. Non-reconstructive aggregate results may be retained.
If MicroAGI removes or replaces a clip, stop using and delete that version after notice. Refer anyone who believes they appear in the data to data@micro-agi.com; do not investigate.
Access grants no ownership, commercialisation, publicity, trademark, or other rights beyond the limited permitted use above.
You are responsible for applicable law and institutional approvals. Access terminates upon breach; on termination, stop use and delete the data.
The data is supplied “AS IS,” without warranties to the extent permitted by law. To that extent, MicroAGI is not liable for indirect, incidental, special, consequential, or exemplary loss arising from its use.
Any governing-law or venue term required for a recipient must be agreed in writing by MicroAGI before access is granted.

The canonical downloadable copy is DATA-USE-AGREEMENT (version 2026-08-28), which is the operative agreement presented to access requesters.

Reporting numbers

Name the suite and its version (phone_edge_cases, this repo revision) and say which categories you sliced on. Pin the revision — hf://datasets/MicroAGI-Labs/hand-tracking-evals@<sha>/... — or the number is not reproducible.

Downloads last month: 29

Total file size:

587 MB