load_rehab_pile_dataset

load_rehab_pile_dataset(name: str, split: Literal['train', 'test'] = 'train', fold: int = 0, extract_path: str | Path | None = None, return_meta: bool = False) → tuple[ndarray, ndarray] | tuple[ndarray, ndarray, dict[str, Any]]

Load a dataset from the RehabPile collection.

The RehabPile collection, introduced in [1], is a unified archive of existing rehabilitation datasets designed to serve as a benchmark for automated human motion assessment. Unlike general human activity recognition, the goal of rehabilitation assessment is to analyze the quality of movement within the same action class, requiring the detection of subtle deviations from an ideal motion.

The collection is composed of 8 original repositories, compiled into 39 classification and 21 regression problems. The data consists of skeleton-based human motion sequences captured from video streams or inertial sensors. Each problem is divided into multiple cross-validation folds to ensure robust evaluation.

Parameters:

namestr

The name of the RehabPile dataset, e.g., “KIMORE_clf_bn_TR”. A full list can be obtained from load_rehab_pile_classification_datasets() or load_rehab_pile_regression_datasets().

split{“train”, “test”}, default=”train”

The split of the data to return.

foldint, default=0

The cross-validation fold (resample) to load. Defaults to the first fold (0). The number of available folds varies by dataset.

extract_pathstr or Path, default=None

The path to look for the data. If no path is provided, the function looks in aeon/datasets/local_data/.

return_metabool, default=False

If True, returns a tuple (X, y, meta_data). The meta_data is a dictionary containing information about the dataset, loaded from the info.json file.

Returns:

Xnp.ndarray

The time series data of shape (n_cases, n_channels, n_timepoints).

ynp.ndarray

The target values (class labels or regression values) of shape (n_cases,).

meta_datadict, optional

A dictionary containing metadata for the dataset, such as the number of classes, channels, and series length. Only returned if return_meta is True.

Notes

webpage: https://msd-irimas.github.io/pages/DeepRehabPile/#rehab-pile

References

[1]

Ismail-Fawaz, Ali, Maxime Devanne, Stefano Berretti, Jonathan Weber, and Germain Forestier.”Deep Learning for Skeleton Based Human Motion Rehabilitation Assessment:A Benchmark.” arXiv preprint arXiv:2507.21018 (2025).