# 3rd WEAR Dataset Challenge @HASCA 2026

![WEAR multimodal activity recognition research illustration](assets/header-v2.png)

Leakage-aware research scaffold for classifying 19 human activities from
one-second wearable inertial signals and pre-extracted egocentric VideoMAE
features. The project prioritizes participant-grouped validation and
reproducible evidence over unverified leaderboard claims.

> Status: research pipeline and hybrid feature infrastructure are available.
> A full trainer, complete official dataset, OOF score, and Kaggle submission
> are not included in this release.

## What this release contains

- Read-only schema checks and deterministic sample profiling.
- A train-window index retaining participant/run groups and label quality.
- IMU and VideoMAE summary features with a quality-aware probability-fusion
  helper.
- Group-disjoint fold assignment, leakage assertions, and the 19-class
  macro-F1 implementation.
- An append-only experiment ledger that records route, billing, tokens,
  latency, OOF scores, hardware time, and JEV decisions.
- A JEV-directed experiment protocol; the router remains independent from the
  model runtime.

![Hybrid IMU and VideoMAE pipeline](assets/hybrid-architecture-v2.png)

## Data and challenge metric

The challenge evaluates 19 activity classes with macro-F1. Training contains
continuous participant recordings with four sensor locations; test provides
one sensor location per one-second window together with 15 VideoMAE feature
vectors. The complete challenge dataset is distributed separately and is not
included here. Obtain it from the [official Kaggle data page](https://www.kaggle.com/competitions/3rd-wear-dataset-challenge-hasca-2026/data)
and follow its current license and rules.

The bundled sample data used during development is intentionally excluded from
this package. Do not commit full data, credentials, or hidden-test predictions.

## Install and inspect

Python 3.11 or later is required.

```powershell
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e ".[dev]"
```

For a local Kaggle download stored outside the repository:

```powershell
python src/inspect_schema.py --data-root E:/data/wear
python src/profile_dataset.py `
  --train-csv E:/data/wear/train/inertial_feat/sbj_0.csv `
  --video E:/data/wear/train/videomae_feat/sbj_0.npy `
  --output artifacts/sbj_0_profile.json
python src/build_train_index.py `
  --train-root E:/data/wear/train `
  --output artifacts/train_index.csv
```

Inspect the actual array shapes before constructing any model input. Keep the
test row order unchanged.

## Validation and experiment loop

Every split must keep each participant and repeated recording together, and
must happen before overlapping windows are generated. Fit normalization,
feature selection, and calibration only on the training groups in each fold.
Compare IMU-only, video-only, probability fusion, and quality-gated fusion on
the same grouped OOF folds.

![Participant-grouped out-of-fold validation](assets/subject-oof-validation-v2.png)

JEV chooses each next hypothesis, model route, and pre-run acceptance
criterion. The local tools execute that choice and save evidence. The
architecture does not automatically accept a score or submit to Kaggle.

![JEV-directed reproducible experiment cycle](assets/jev-experiment-loop-v2.png)

See [the experiment protocol](docs/EXPERIMENT_PROTOCOL.md) and
[the JEV-directed architecture](docs/COGNITIVE_ARCHITECTURE.md).

## Tests

```powershell
$env:PYTEST_DISABLE_PLUGIN_AUTOLOAD = "1"
python -m pytest -q --basetemp .pytest-tmp
python -m compileall -q src tests
```

The environment flag avoids unrelated globally installed pytest plugins from
being auto-loaded. Test results do not constitute a Kaggle score.

## Rules, licensing, and scope

- This code is MIT licensed; see [LICENSE](LICENSE).
- Dataset terms are separate and apply to the data, not this code. The full
  dataset is not redistributed here.
- Check [the recorded rules](docs/RULES.md) and current official Kaggle rules
  before any submission or use of external data/weights.
- No rank-1 claim is made. A competitive claim requires an official leaderboard
  result and reproducible evaluation evidence.

## References

- [Official challenge page](https://mariusbock.github.io/wear/challenge.html)
- [Kaggle competition and data](https://www.kaggle.com/competitions/3rd-wear-dataset-challenge-hasca-2026)
- WEAR dataset paper: [doi:10.1145/3699776](https://doi.org/10.1145/3699776)
- [Visual asset prompts and provenance](docs/ASSET_GENERATION.md)