Rehab-Pile: Unified Rehab Motion Benchmark
- Rehab-Pile is a unified benchmark for automated rehabilitation motion assessment, emphasizing intra-exercise quality analysis using skeletal data.
- It aggregates 9 public repositories into 60 standardized datasets for both classification and regression tasks while preserving original joint configurations.
- The framework enforces reproducible evaluation with consistent preprocessing and cross-subject splits to address fragmented methodologies in motion assessment.
Rehab-Pile is a unified benchmark and open archive for automated rehabilitation motion assessment built from skeleton sequences extracted from video or depth cameras. Its central purpose is to standardize a research area in which the target is not generic action recognition, but quality-of-movement analysis within the same exercise, including binary and multi-class quality/error detection and extrinsic regression of continuous quality labels or clinical severity. The benchmark addresses fragmented datasets, inconsistent splits and preprocessing, and hard-to-reproduce pipelines by aggregating 9 public repositories into 60 standardized datasets—39 classification datasets and 21 extrinsic regression datasets—under a single reproducible framework with released datasets, source code, and results (Ismail-Fawaz et al., 28 Jul 2025).
1. Definition and problem setting
Rehab-Pile is organized around a distinction that is fundamental in rehabilitation AI but often obscured by the vocabulary of general human activity recognition. In general HAR, the task is to distinguish among different actions such as walking and running. In Rehab-Pile, the task is to assess subtle deviations from an “ideal” execution of the same rehabilitation exercise, or to predict a continuous external score attached to performance quality (Ismail-Fawaz et al., 28 Jul 2025).
This distinction determines the structure of the archive. Each exercise in an original repository is treated as a separate dataset, because assessing multiple exercises with one model is considered impractical in the benchmark design. The result is a per-exercise evaluation regime rather than a single monolithic multi-exercise corpus. The benchmark therefore foregrounds exercise-specific motion quality, cross-subject generalization, and reproducible preprocessing over broad action taxonomy coverage (Ismail-Fawaz et al., 28 Jul 2025).
A common misconception is that rehabilitation assessment can be evaluated with the same assumptions as generic pose or action benchmarks. Rehab-Pile formalizes the opposite view: the relevant signal is intra-class movement quality, not inter-class action identity. This emphasis aligns with rehabilitation-specific dataset work showing severe domain shift when models trained on generic depth data are transferred to rehabilitation movements, especially when pediatric subjects, limited range of motion, self-contact, and clinically atypical postures are present (Banik et al., 2021).
2. Corpus composition and dataset families
Rehab-Pile aggregates repositories spanning Kinect skeletons, Kinect v2 plus Vicon, and OpenPose-derived 2D skeletons. The available summary specifies the following repository families and their task roles (Ismail-Fawaz et al., 28 Jul 2025).
| Repository family | Task types | Representative details |
|---|---|---|
| EHE | Extrinsic regression | 6 datasets; Alzheimer’s severity in |
| KIMORE | Regression and binary classification | 5 exercises; Kinect; score $0$–$100$ and binary thresholding |
| UI-PRMD | Regression and binary classification | 10 exercises; Kinect; continuous label in |
| IRDS | Binary classification | 9 datasets; Kinect; healthy and patient cohorts |
| KERAAL | Binary and multi-class classification | 3 datasets each for BN and MC; Kinect v2 + Vicon |
| KINECAL | Binary classification | 4 datasets; at-risk vs not |
| SPHERE | Binary classification | 1 dataset; anomaly detection reformulated as binary classification |
| UCDHE | Binary and multi-class classification | 2 datasets each; OpenPose 2D skeletons |
The benchmark states that 9 public repositories are aggregated into 60 standardized datasets. In the available technical summary, the repository families above are explicitly detailed and account for the reported 21 regression datasets and 39 classification datasets (Ismail-Fawaz et al., 28 Jul 2025).
The datasets are heterogeneous in subject count, sequence length, joint count, coordinate dimensionality, and label semantics. EHE uses Kinect skeletons with 25 joints and 8 dimensions per joint; KIMORE uses 18 joints and 3 dimensions; UI-PRMD uses 22 joints and 3 dimensions; KERAAL uses 11 joints and 7 dimensions; UCDHE uses OpenPose 2D skeletons with 8 joints and 2 dimensions. This heterogeneity is not erased by aggressive canonicalization. Instead, Rehab-Pile preserves original joint layouts, original coordinate types, and dataset-specific adjacency when graph models such as STGCN are used (Ismail-Fawaz et al., 28 Jul 2025).
This design has methodological consequences. The benchmark does not force all datasets into a single anatomical schema, and it does not redefine joint ordering or bone connectivity globally. A plausible implication is that Rehab-Pile privileges faithful archival standardization over anatomical homogenization, thereby making cross-dataset comparisons reproducible without pretending that all upstream capture pipelines are equivalent.
3. Representation, preprocessing, and data standardization
The core data object in Rehab-Pile is a skeleton sequence represented as
where is the number of frames, the number of joints, and the number of features per joint. With the exception of STGCN, all models consume a multivariate time series obtained by flattening the joint and feature axes to shape . STGCN alone keeps the joint axis and graph structure explicit (Ismail-Fawaz et al., 28 Jul 2025).
Temporal standardization is performed within each dataset by resampling all sequences to the dataset’s average length using Fourier-based resampling via scipy.signal.resample. This assumes a uniform sampling rate within each dataset. Feature scaling is min–max normalization per feature dimension, computed only on the training set and then applied to train and test data:
where $0$0 and $0$1 are the training-set minimum and maximum for dimension $0$2 (Ismail-Fawaz et al., 28 Jul 2025).
The benchmark is deliberately conservative about additional preprocessing. No limb-length normalization, root-centering, or engineered kinematics such as angles, velocities, or jerk are introduced by the benchmark itself. Missing joints, occlusions, and smoothing are not additionally handled beyond the upstream skeleton estimation provided by the original repositories. Raw per-joint features are therefore the standardized input substrate (Ismail-Fawaz et al., 28 Jul 2025).
This choice is technically important. It prevents leakage from handcrafted harmonization steps and makes model comparison cleaner, but it also means that upstream sensor idiosyncrasies remain part of the learning problem. That trade-off is consistent with Rehab-Pile’s role as a benchmark archive rather than a fully re-authored multimodal dataset.
4. Tasks, splits, and benchmarking protocol
Rehab-Pile supports two main task families. The first is classification, including binary correctness detection and multi-class error-type recognition. The second is extrinsic regression, in which a model predicts a continuous external score such as a clinical severity or quality assessment. Examples include KIMORE regression scores in $0$3, EHE labels in $0$4, and UI-PRMD labels in $0$5 (Ismail-Fawaz et al., 28 Jul 2025).
Evaluation protocols are standardized around cross-subject splits. If healthy and unhealthy subjects are both present, the benchmark uses 5-fold cross-subject evaluation when there are at least 10 unhealthy subjects, and leave-one-unhealthy-subject-out when there are fewer than 10. If only healthy subjects are present, the protocol is 5-fold when there are at least 10 subjects and leave-one-subject-out otherwise. When applicable, tests include only unhealthy subjects. KERAAL is treated as a special case: all healthy subjects are placed in training, and patient sequences are evaluated with leave-one-subject-out (Ismail-Fawaz et al., 28 Jul 2025).
The training and reporting protocol also includes repeated random initialization. For each fold, the same model is trained five times with different random seeds and predictions are averaged at test time. Results are then averaged across folds. Implementations are unified in a single TensorFlow/Keras codebase with identical training loops across models (Ismail-Fawaz et al., 28 Jul 2025).
The benchmark paper frames this as a response to a field-wide reproducibility problem. Because prior rehabilitation-motion studies frequently used incompatible splits and preprocessing conventions, performance numbers were difficult to compare directly. Rehab-Pile converts that fragmented landscape into a controlled per-exercise benchmarking framework. This suggests that its primary contribution is infrastructural as much as algorithmic.
5. Relation to rehabilitation-specific data and domain shift
The rationale for a benchmark like Rehab-Pile is reinforced by rehabilitation-specific pose estimation results outside the benchmark itself. The Vogtareuth Rehab Depth Datasets show that a pose model trained on the non-rehab ITOP depth benchmark achieves PCK 89.21% and PCKh 89.12% on ITOP, but when evaluated on the rehabilitation-specific VtR dataset at $0$6, performance drops to PCK 41.08% and PCKh 22.31%. Fine-tuning on VtR raises performance on the VtR test set to PCK 93.90% and PCKh 86.03% (Banik et al., 2021).
Those results quantify a general point that underlies Rehab-Pile: rehabilitation movements are not just another action-recognition domain. Pediatric subjects, complex postures, limited mobility, self-contact, and atypical clinical environments create a domain shift that generic benchmarks do not capture (Banik et al., 2021). Rehab-Pile operationalizes that insight by centering rehabilitation repositories and enforcing exercise-specific evaluation rather than relying on transfer from generic motion datasets (Ismail-Fawaz et al., 28 Jul 2025).
A related example is Keraal, which is included in Rehab-Pile as binary and multi-class classification datasets. In its standalone clinical presentation, Keraal is a low-back pain rehabilitation dataset recorded in a double-blind clinical study, with annotations for motion assessment, error recognition, spatial localization, and temporal localization (Nguyen, 5 Jan 2026). Rehab-Pile uses KERAAL for Challenge 1 and Challenge 2 style classification, but the broader Keraal annotation design also points to future benchmark directions beyond the current classification/regression core.
6. Significance, reproducibility, and limitations
Rehab-Pile matters because it changes the unit of comparison in rehabilitation AI. Instead of isolated papers with incompatible preprocessing, it offers a unified archive, a general benchmarking framework, and extensive benchmarking of multiple architectures across classification and regression tasks (Ismail-Fawaz et al., 28 Jul 2025). For researchers, this means that claims about model superiority can be tested under matched splits, matched normalization, and matched sequence-length handling.
Its reproducibility claims are unusually explicit for the domain. The paper states that the datasets, source code, and results are all publicly available, and the archive preserves instructions to obtain the underlying public data (Ismail-Fawaz et al., 28 Jul 2025). That directly addresses a recurring obstacle in rehabilitation assessment research, where methodological novelty often outpaces reproducible evaluation infrastructure.
The benchmark also has clear boundaries. It preserves original dataset heterogeneity rather than imposing a universal skeletal ontology. It does not add denoising, occlusion repair, or engineered biomechanical features. It treats each exercise as a separate dataset rather than pursuing a unified multi-exercise model. These are not defects so much as defining choices. A plausible implication is that Rehab-Pile should be understood as a foundational benchmark layer onto which richer formulations—such as temporal error localization, multimodal sensing, or clinician-in-the-loop intelligent tutoring system tasks—may later be added.
In that sense, Rehab-Pile is best viewed as both an archive and a methodological baseline for rehabilitation motion assessment. It standardizes the evaluation substrate for skeleton-based assessment while leaving open the next research questions: how to model temporal error structure, how to combine 2D, 3D, and clinical metadata, and how to move from per-exercise benchmarking toward clinically actionable assessment pipelines (Ismail-Fawaz et al., 28 Jul 2025).