MonicaVahd/skeletoneqa-weights-nc

SkeletonEQA weights (non-commercial): checkpoints trained on REHAB24-6, and all CTR-GCN checkpoints

Weights only, no dataset samples. These weights contain no dataset samples; obtain the data from the original authors. They are the model checkpoints of the benchmark study Skeleton-Based Exercise Quality Assessment Under Domain Shift (MCHM '26, DOI 10.1145/3840471.3842592), a binary squat-quality classifier (0 = incorrect, 1 = correct) on 18-joint OpenPose skeleton sequences of shape (N, 3, 100, 18, 1). This repository holds 368 files, 3.18 GB.

What the files are and how they are named

• Every file is a PyTorch state dict (a flat OrderedDict of tensors; no optimizer state, no wrapper keys). Layout checked by reading the zip listing and pickle opcodes of the public-condition files without loading them.
• Path: <model>/<folder>/<file>.pth, with <model> one of the lower-case model names and <folder> the experimental condition (table below).
• File names inside the folders keep the names used during training: <model>_ood_b<n>_<train>_to_<test>.pth, <model>_ood_c<n>_..._to_<test>.pth, <model>_loso_<dataset>_round<k>.pth, <model>_kfold_custom_fold<k>.pth, <model>_mixed_all_datasets.pth; ST-GCN files have no <model>_ prefix. A suffix _rerun marks the second training run of a condition (see "Preprocessing").
• MANIFEST.csv lists every file with size, SHA-256, model, condition, training pool, fold and repository.
• A folder name ending in _loso_run2 holds a second set of LOSO fold files written on 2026-06-21 and 2026-06-22; D1_ablation_* holds a normalisation ablation on the pooled data; B3_supcon holds a SupCon variant of B3. These three groups have no table in the paper.

Code, training and seeds

• Code: https://github.com/MonicaVahd/FAIRCoachOOD, tag v0.1.0, commit 859324b343a2bf4fbd25376004692d9da078f9c6.
• Training protocol (paper section 3.3): Adam, learning rate 1e-3, weight decay 1e-4, ReduceLROnPlateau (patience 10, factor 0.5), batch size 16, at most 200 epochs, early stopping on validation loss (patience 20), inverse-frequency weighted cross-entropy. The stored SkateFormer D1 script uses an early-stopping patience of 50 (docs/KNOWN_DIFFERENCES.md, D5).
• Seed 42, single run per file. Re-training is not expected to reproduce a file exactly.

Conditions in this repository

Models in this repository: ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer. Training pool and test target follow Table 3 of the paper (B: 80 % of one dataset; C: 90 % pool of two or three datasets; A: leave-one-subject-out or 5-fold). Folders whose pool contains Custom (A3_kfold, B5, B6, B12, C2, C3, C6, C7, C9, C10, D1_ablation_hip_centering_only, D1_ablation_hip_centering_torso_scaling, D1_ablation_no_normalization, D1_mixed): the weights were trained partly on publicly available instructional videos that are not redistributed.

Folder	Protocol	Training pool	Test target	Models	Files	File dates	Paper table (by name and date only)
A1_loso	LOSO	EC3D	held-out EC3D subject	CTR-GCN	4	2026-07-14 .. 2026-07-14	Table 5
A1_loso_run2	LOSO, second file set	EC3D	held-out EC3D subject	CTR-GCN	4	2026-06-21 .. 2026-06-21	none (second LOSO file set)
A2_loso	LOSO	REHAB24-6	held-out REHAB24-6 subject	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	81	2026-07-14 .. 2026-07-15	Table 5
A2_loso_run2	LOSO, second file set	REHAB24-6	held-out REHAB24-6 subject	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	81	2026-06-21 .. 2026-06-21	none (second LOSO file set)
A3_kfold	5-fold CV	Custom	held-out Custom fold	CTR-GCN	5	2026-07-14 .. 2026-07-14	Table 5
A4_loso	LOSO	UI-PRMD	held-out UI-PRMD subject	CTR-GCN	10	2026-07-14 .. 2026-07-14	Table 5
A4_loso_run2	LOSO, second file set	UI-PRMD	held-out UI-PRMD subject	CTR-GCN	10	2026-06-21 .. 2026-06-21	none (second LOSO file set)
B1	single-source OOD	EC3D	REHAB24-6	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
B2	single-source OOD	EC3D	Custom	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
B3	single-source OOD	REHAB24-6	EC3D	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-09 .. 2026-06-18	Table 6
B3_supcon	single-source OOD, SupCon variant	REHAB24-6	EC3D	ST-GCN	1	2026-07-03 .. 2026-07-03	none (SupCon variant)
B4	single-source OOD	REHAB24-6	Custom	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-09 .. 2026-06-18	Table 6
B5	single-source OOD	Custom	EC3D	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
B6	single-source OOD	Custom	REHAB24-6	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
B7	single-source OOD	UI-PRMD	EC3D	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
B8	single-source OOD	UI-PRMD	REHAB24-6	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
B9	single-source OOD	UI-PRMD	Custom	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
B10	single-source OOD	EC3D	UI-PRMD	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
B11	single-source OOD	REHAB24-6	UI-PRMD	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-09 .. 2026-06-18	Table 6
B12	single-source OOD	Custom	UI-PRMD	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
C1	multi-source OOD	EC3D + REHAB24-6	Custom	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-15 .. 2026-06-19	Table 6
C2	multi-source OOD	EC3D + Custom	REHAB24-6	CTR-GCN	1	2026-06-16 .. 2026-06-16	Table 6
C3	multi-source OOD	REHAB24-6 + Custom	EC3D	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-15 .. 2026-06-19	Table 6
C4	multi-source OOD	EC3D + UI-PRMD	Custom	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
C5	multi-source OOD	REHAB24-6 + UI-PRMD	Custom	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	18	2026-06-15 .. 2026-06-22	Table 6
C6	multi-source OOD	EC3D + REHAB24-6 + Custom	UI-PRMD	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-15 .. 2026-06-19	Table 6
C7	multi-source OOD	Custom + UI-PRMD	EC3D	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
C8	multi-source OOD	EC3D + REHAB24-6 + UI-PRMD	Custom	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	18	2026-06-15 .. 2026-06-22	Table 6
C9	multi-source OOD	EC3D + Custom + UI-PRMD	REHAB24-6	CTR-GCN	2	2026-06-16 .. 2026-06-22	Table 6
C10	multi-source OOD	REHAB24-6 + Custom + UI-PRMD	EC3D	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	18	2026-06-15 .. 2026-06-22	Table 6
D1_ablation_hip_centering_only	pooled 80/10/10, normalisation ablation	EC3D + REHAB24-6 + Custom + UI-PRMD	held-out 10 % of the pooled data	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-20 .. 2026-06-20	none (normalisation ablation)
D1_ablation_hip_centering_torso_scaling	pooled 80/10/10, normalisation ablation	EC3D + REHAB24-6 + Custom + UI-PRMD	held-out 10 % of the pooled data	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-20 .. 2026-06-20	none (normalisation ablation)
D1_ablation_no_normalization	pooled 80/10/10, normalisation ablation	EC3D + REHAB24-6 + Custom + UI-PRMD	held-out 10 % of the pooled data	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	9	2026-06-20 .. 2026-06-20	none (normalisation ablation)
D1_mixed	pooled 80/10/10	EC3D + REHAB24-6 + Custom + UI-PRMD	held-out 10 % of the pooled data	ST-GCN, CTR-GCN, InfoGCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer, HyperFormer, FreqMixFormer	18	2026-06-22 .. 2026-07-15	Table 5 (D1)

Preprocessing

All sequences were converted to the 18-joint OpenPose layout, hip-centred (right hip), torso-scaled, padded/trimmed (UI-PRMD: interpolated) to 100 frames, transposed to (3, 100, 18), min-max normalised in x and y, and given a person axis, as in section 3.2 of the paper; the converters are in the code repository.

UI-PRMD normalize_xy note (code repository, docs/KNOWN_DIFFERENCES.md, D3 and D4).

• The UI-PRMD re-run script states that it followed a preprocessing fix in which normalize_xy is applied to UI-PRMD.
• Files whose training pool contains UI-PRMD exist in two versions for conditions B7, B8, B9, C4, C5, C7, C8, C9, C10 and D1: the original files (written 2026-06-15 .. 06-19; D1: 2026-07-15) and files with the suffix _rerun (written 2026-06-22). Both are included. B10, B11, B12 and C6 have a single file (their original checkpoints were re-evaluated on refreshed test data in the re-run).
• Which UI-PRMD preprocessing each original file was trained with is not recorded. The paper's section 6.5 reports 25 status changes between the two sets (11 recoveries, 8 regressions, 6 FC<->PC); the code repository reproduces that count from the stored confusion counts (D4).

Limitations and provenance

• The results are exploratory: one run per condition, seed 42, and the paper reports no confidence intervals.
• 77.8 % of the 198 OOD experiments collapse (154 of 198: 141 full collapse, 13 partial collapse): the classifier predicts one class for (nearly) all test samples. A non-zero macro-F1 of a collapsed cell reflects majority-class prediction, not transfer. Many of these checkpoints are therefore weak classifiers outside their training distribution.
• Files were matched to the paper only by name and date. No hash, log or manifest ties a file to a value in the paper's tables; if a file with the same name was overwritten, the statement would not detect it.
• The paper's cell values (macro-F1 and FC/PC/OK) are in results/ood_results.csv of the code repository (Table 6) and results/iid_results, results/mixed_results (Table 5).

Licence

This repository is licensed CC BY-NC 4.0 (license: cc-by-nc-4.0): the most restrictive licence that applies to its contents.

• REHAB24-6 (training data of every REHAB24-6 pool): Zenodo record licence field "Creative Commons Attribution Non Commercial 4.0 International"; terms: "This dataset is for academic or non-profit organization noncomercial research use only." and "By using you agree to appropriately reference the paper above in any publication making of its use."
• CTR-GCN (all CTR-GCN files, whatever their pool): upstream code licence CC BY-NC 4.0.

Other components, for the pools and models in this repository:

Component	Licence / status
EC3D	EC3D licence not stated
UI-PRMD	Open Data Commons Public Domain Dedication and License v1.0 (dataset paper footnotes)
Custom (every folder whose training pool contains it, see "Conditions in this repository"; includes the CTR-GCN folders with Custom in the pool)	trained partly on publicly available instructional videos that are not redistributed
ST-GCN code	BSD-2-Clause
HD-GCN, ProtoGCN, SkateFormer code	MIT
BlockGCN, HyperFormer code (modified in the code repository)	Apache-2.0
CTR-GCN code	CC BY-NC 4.0
InfoGCN code	no licence file in the upstream repository
FreqMixFormer code	no licence file in the upstream repository

InfoGCN and FreqMixFormer: the files are included, but the upstream code of both models has no licence file and is not distributed in the code repository. It must be fetched with scripts/fetch_upstream.py; the released builders do not load these two models. The CTR-GCN upstream code must be fetched the same way.

Loading

Tested with the released builders of the code repository at tag v0.1.0 (PyTorch 2.4.1, CPU): one checkpoint per model was loaded with torch.load(..., weights_only=True) into the builder's untrained model with strict=True, followed by a forward pass on a random input (4, 3, 100, 18, 1), which returned logits of shape (4, 2) for ST-GCN, BlockGCN, HD-GCN, ProtoGCN, SkateFormer and HyperFormer (the test used the B1 files of the open repository). Not tested: CTR-GCN, InfoGCN, FreqMixFormer (their builders are not part of the release).

import sys, torch
sys.path.insert(0, "FAIRCoachOOD")                      # checkout of the code repository at tag v0.1.0
from models.builders import build_hdgcn_model           # also: build_stgcn_model, build_blockgcn_model,
                                                        #       build_protogcn_model, build_skateformer_model, build_hyperformer_model
model = build_hdgcn_model(device="cpu")                 # untrained model
state = torch.load("hdgcn/B3/hdgcn_ood_b3_rehab24_to_ec3d.pth", map_location="cpu", weights_only=True)
model.load_state_dict(state, strict=True)
model.eval()
logits = model(torch.rand(4, 3, 100, 18, 1))            # (4, 2); class 1 = correct squat

Forward signatures: ST-GCN, HD-GCN and ProtoGCN model(x); SkateFormer model(x, index_t) with index_t = zeros(N, 100, dtype=long); BlockGCN model(x, y_onehot, x) (returns logits and an auxiliary loss); HyperFormer model(x, y) (returns logits and labels). build_*_model(checkpoint_path=...) in the code repository also loads a file, but it calls torch.load without weights_only=True; the recipe above is the safer variant.

References

Paper

@inproceedings{10.1145/3840471.3842592,
  author    = {Vahdati, Monica and Gholizadeh HamlAbadi, Kamran and Etemad, Ali and El Saddik, Abdulmotaleb},
  title     = {Skeleton-Based Exercise Quality Assessment Under Domain Shift: A Benchmark Study of Cross-Domain Failure Modes},
  year      = {2026},
  isbn      = {9798400729256},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  url       = {https://doi.org/10.1145/3840471.3842592},
  doi       = {10.1145/3840471.3842592},
  booktitle = {Proceedings of the 3rd International Workshop on Multimedia Computing for Health and Medicine},
  pages     = {43--52},
  numpages  = {10},
  location  = {Rio de Janeiro, Brazil},
  series    = {MCHM '26}
}

Project page: https://skeletoneqa.github.io/. Code: https://github.com/MonicaVahd/FAIRCoachOOD (tag v0.1.0).

Datasets

• EC3D: Ziyi Zhao, Sena Kiciroglu, Hugues Vinzant, Yuan Cheng, Isinsu Katircioglu, Mathieu Salzmann, Pascal Fua. 3D pose based feedback for physical exercises. Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1316-1332. arXiv:2208.03257 (https://arxiv.org/abs/2208.03257). Dataset repository: https://github.com/Jacoo-Zhao/3D-Pose-Based-Feedback-For-Physical-Exercises.
• REHAB24-6: Andrej Černek, Jan Sedmidubsky, Petra Budikova. REHAB24-6: Physical therapy dataset for analyzing pose estimation methods. International Conference on Similarity Search and Applications (SISAP), Springer, 2024, pp. 18-33. Zenodo record: https://zenodo.org/records/13305826.
• UI-PRMD: Aleksandar Vakanski, Hyung-pil Jun, David Paul, Russell Baker. A data set of human body movements for physical rehabilitation exercises. Data 3(1):2, 2018. https://pmc.ncbi.nlm.nih.gov/articles/PMC5773117/. Project page: https://webpages.uidaho.edu/ui-prmd/.

Model architectures (upstream papers and repositories)

• ST-GCN: Sijie Yan, Yuanjun Xiong, Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. AAAI 2018. https://github.com/yysijie/st-gcn
• CTR-GCN: Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. ICCV 2021, pp. 13359-13368. https://github.com/Uason-Chen/CTR-GCN
• InfoGCN: Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, Karthik Ramani. InfoGCN: Representation learning for human skeleton-based action recognition. CVPR 2022, pp. 20186-20196. https://github.com/stnoah1/infogcn
• BlockGCN: Yuxuan Zhou, Xudong Yan, Zhi-Qi Cheng, Yan Yan, Qi Dai, Xian-Sheng Hua. BlockGCN: Redefine topology awareness for skeleton-based action recognition. CVPR 2024, pp. 2049-2058. https://github.com/ZhouYuxuanYX/BlockGCN
• HD-GCN: Jungho Lee, Minhyeok Lee, Dogyoon Lee, Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. ICCV 2023, pp. 10444-10453. https://github.com/Jho-Yonsei/HD-GCN
• ProtoGCN: Hongda Liu, Yunfan Liu, Min Ren, Hao Wang, Yunlong Wang, Zhenan Sun. Revealing key details to see differences: A novel prototypical perspective for skeleton-based action recognition. CVPR 2025, pp. 29248-29257. https://github.com/firework8/ProtoGCN
• SkateFormer: Jeonghyeok Do, Munchurl Kim. SkateFormer: Skeletal-temporal transformer for human action recognition. ECCV 2024, pp. 401-420. https://github.com/KAIST-VICLab/SkateFormer
• HyperFormer: Yuxuan Zhou, Zhi-Qi Cheng, Chao Li, Yanwen Fang, Yifeng Geng, Xuansong Xie, Margret Keuper. Hypergraph transformer for skeleton-based action recognition. arXiv:2211.09590, 2022. https://github.com/ZhouYuxuanYX/Hyperformer
• FreqMixFormer: Wenhan Wu, Ce Zheng, Zihao Yang, Chen Chen, Srijan Das, Aidong Lu. Frequency guidance matters: Skeletal action recognition by frequency-aware mixed transformer. ACM Multimedia 2024, pp. 4660-4669. https://github.com/wenhanwu95/FreqMixFormer