# 3rd WEAR Dataset Challenge @HASCA 2026

![WEAR header](assets/header.png)

An open-source starter for the [3rd WEAR Dataset Challenge @HASCA 2026](https://www.kaggle.com/competitions/3rd-wear-dataset-challenge-hasca-2026). It combines one-second inertial windows with pre-extracted VideoMAE features to classify 19 human-activity labels using participant-aware validation.

![Multimodal fusion](assets/multimodal-fusion.png)

## Data contract

- Training data: participant-level CSV streams at 50 Hz plus VideoMAE features at 30 FPS.
- Test data: 12,234 pre-windowed examples with inertial, video, and metadata arrays.
- Targets: `null` plus 18 activities.
- Metric: macro-F1.

The local `data/sample/` directory contains only a small schema contract. The full dataset is approximately 8 GB and is intentionally not versioned.

![Subject-aware validation](assets/subject-aware-validation.png)

## Reproduce

```bash
python src/inspect_schema.py --data-root data/sample
python src/build_train_index.py --train-root /path/to/train --output artifacts/train_index.csv
python src/audit_submission.py data/sample/sample_submission.csv --sample data/sample/sample_submission.csv
```

Keep participants disjoint between training and validation. Preserve official test ordering and audit exact submission columns before upload.

![Temporal fusion](assets/temporal-fusion.png)

## Scope and licensing

This repository is an engineering baseline, not a clinical activity-recognition device or a claim of human-performance validity. Dataset terms are CC BY-NC-SA 4.0 as stated by Kaggle; do not redistribute the full data. Original code is MIT-licensed in [LICENSE](LICENSE).

## References

- [Challenge page](https://www.kaggle.com/competitions/3rd-wear-dataset-challenge-hasca-2026)
- [WEAR challenge specification](https://mariusbock.github.io/wear/challenge.html)
- [WEAR paper](https://doi.org/10.1145/3699776)