SCAI-SCI Gait Dataset

Multi-Modal Clinical Benchmark of Walking in Spinal Cord Injury: Optical Motion Capture, Ground Reaction Forces, Surface EMG, and Video-Derived SMPL 3D Body Models

---

Important

The data are not hosted in this repository. The optical motion capture trials, SMPL body model sequences, 3D renders, and clinical metadata contain sensitive patient information. They are distributed through PhysioNet as credentialed-access data, under a Data Use Agreement that strictly prohibits any attempt to identify or re-identify a participant.

This repository provides the official documentation, complete data dictionary, interactive 3D viewers, data loading examples, and the processing pipeline that produced the SMPL sequences.

---

Overview

Instrumented 3D gait analysis is the clinical gold standard for assessing walking after spinal cord injury (SCI). However, it requires a specialized motion-capture laboratory that few clinical centers possess. Markerless video methods could bring accessible gait assessment to broader clinical and home settings, but they are rarely validated on neurological gait.

Schematic illustration of the dataset (not participant data): sagittal walking video reconstructed as a 3D SMPL body model, with laboratory motion capture and force plates as the reference.

The SCAI-SCI Gait Dataset bridges this gap by pairing both modalities in a comprehensive clinical release. It covers 240 adults with traumatic or non-traumatic SCI, assessed in the specialized gait laboratory of the Swiss Paraplegic Centre (Nottwil, Switzerland) during routine clinical care. For each participant, it provides synchronized optoelectronic motion capture with ground reaction forces from two force plates (and surface EMG in 194 trials), 3D SMPL body models reconstructed with CameraHMR from sagittal-plane video of the same session, and a clinical table standardized to the International SCI Core Data Set v3.0.

	Count
Participants	240
Motion-capture trials (.c3d)	300 (180 participants with 1 trial, 60 with 2)
Marker frames	222,400 (23 min of walking)
Trials with surface EMG	194 (171 participants)
SMPL sequences (.pkl)	300 (239 participants)
SMPL mesh frames	44,980 (at 15 fps)
SMPL renders (.mp4)	299
Clinical table	240 rows × 14 columns
Per-trial gait parameters	300 trials (292 with a full gait cycle)

---

Contents

1. Overview
2. What the release contains
3. Cohort
4. Ethics and de-identification
5. Why CameraHMR
6. This repository
7. Installation
8. Getting the data
9. Viewers
10. Loading the data
11. Known limitations
12. Citation
13. License
14. Acknowledgements
15. Funding
16. Authors and contributions
17. Contact

---

What the release contains

RELEASE_v1.5_SCAI-SCI-Gait/                                            4.3 GB
├── MoCap_Data/      300 .c3d  mocap<studyID>-<trial>.c3d                     209 MB
├── SMPL_Data/       300 .pkl  v<studyID>_sagittal_<li|re>_defaced_smpl_sequence.pkl   3.5 GB
├── SMPL_Videos/     299 .mp4  v<studyID>_sagittal_<li|re>_defaced.mp4        571 MB
├── metadata/
│   ├── medical_metadata.csv   clinical table, 240 × 14
│   ├── data_dictionary.csv    every column of every table and every file field (94 rows)
│   └── cohort_counts_*.csv    6 attributes released only as counts
├── trials.csv       one row per released file (899 rows)
├── gait_parameters.csv  one row per MoCap trial (300 rows)
└── manifest.sha256  SHA-256 of every file (909 entries)

All modalities link via studyID, which forms the prefix of every file name. It is a pseudorandomized number drawn for this release (1–240); it does not correspond to hospital record numbers, nor does it match any other release.

Part	Content	Details
MoCap	20 lower-body markers (IOR protocol) in mm, at 100, 200 or 246 Hz. Two Bertec force plates (6 channels each, C3D type 4). Raw surface EMG in 194 trials.	MoCap_Data/
SMPL sequences	CameraHMR fits of the SMPL body model to each frame of the sagittal video: pose, shape, 24 joint positions and the 6,890-vertex mesh, 46–553 frames per sequence.	SMPL_Data/
SMPL renders	Each sequence rendered as a grey mesh on a black background (H.264, 15 fps). No camera image.	SMPL_Videos/
Clinical table	Sex, age group, neurologic category, level region, completeness, etiology, time since injury, LEMS, SCIM III mobility, walking condition, care setting, vibration sense, spasticity.	metadata/
Index	trials.csv: modality, frames, duration, rates, marker completeness, EMG muscles, force-plate channel map, analog units, SMPL frame range and walking direction.	metadata/data_dictionary.csv
Gait parameters	gait_parameters.csv: strides, walking speed, cadence, stride and step length, step width, stance, swing, double support and stride-time variability, per trial.	MoCap_Data/

The data dictionary in this repository is the one shipped with the release. It is the reference for every column and file field.

---

Cohort

Counts of the 240 participants from the clinical table. Counts below 5 are shown as <5.

Variable	Participants
Sex	male 179; female 61
Age at examination (ISCoS groups)	15–29: 41; 30–44: 61; 45–59: 87; 60–74: 46; 75+: 5
Neurologic category	AIS D, any level: 213; T1–S5 AIS ABC: 18; AIS E (no deficit): 6; C1–4 AIS ABC: <5; not recorded: <5
Region of the neurological level	C1–4: 33; C5–8: 64; T1–S5: 133; intact: 5; not recorded: 5
Completeness (AIS)	incomplete (B–E): 229; complete (A): 9; not recorded: <5
Etiology	traumatic: 171; non-traumatic: 67; not recorded: <5
Years since injury	<1: 143; 1–4: 55; 5–9: 16; 10–14: 5; 15–19: 8; 20–24: <5; 25–29: 8; 30+: <5

Walking speed over the 292 trials with a full gait cycle: median 0.98 m/s (range 0.22–1.53 m/s).

---

Ethics and de-identification

Study design. Data were collected during routine clinical practice and analyzed retrospectively in accordance with the Swiss Human Research Act (HRA, Art. 32–33). The project "Gait in ambulatory patients with spinal cord injury" was approved by the Ethikkommission Nordwest- und Zentralschweiz (EKNZ, BASEC 2022-00935). Amendment 01 (study protocol v3.0), which permits sharing the data under credentialed access, was approved on 29 July 2026. Since 2010, every patient of the Swiss Paraplegic Centre has been informed regarding the potential research use of their routine clinical data and provided the opportunity to decline. Only data from patients who did not decline were included.

Participants. Adults (18 years or older at examination) with traumatic or non-traumatic SCI, capable of walking independently for at least 10 m with or without assistive aids, assessed in the gait laboratory between 2010 and 2022. Six additional participants were excluded prior to release under protocol criteria (5 under 18 years of age at examination, 1 without clinical data).

De-identification followed the SPHN Risk Assessment Template v2.1.2. It was carried out and verified with the SCAI Lab release toolkit:

• Identifiers. Participants carry new pseudorandomized numbers drawn for the release. The key remains on the institution's secure data server. No name, birth date, hospital number, or calendar date appears in any released file.
• C3D headers. Only the POINT, ANALOG, FORCE_PLATFORM, and EVENT groups are retained, with free-text descriptions blanked. Every other group is removed, including the one where acquisition software stores patient fields, operator names, and timestamps. Marker trajectories and analog signals remain uncorrupted.
• Clinical table. Core Data Set variables are reported in standard clinical groups: age in ISCoS groups, time since injury in Core groups, and scores in bands. The neurological level, height, weight, subcategory, etiology codes, and time in rehabilitation are released only as aggregate cohort counts. No per-side level, individual muscle grade, or joint angle is released.
• Video. No raw camera video is released. SMPL is a parametric template mesh without texture or facial geometry, fitted to defaced video footage. The renders display only that mesh on a solid black background.
• Verification. Every file header was audited field by field. The released files were scanned against source identifiers to ensure zero leakage. A file-level re-identification red-team audit was performed, and the formal SPHN risk assessment was scored from the released content.

Because gait dynamics and body shape are biometric and clinical demographics are not k-anonymized, the dataset is accessible exclusively under credentialed access.

---

Why CameraHMR

Five monocular SMPL-based human mesh recovery methods were benchmarked against motion capture on this cohort: CameraHMR, WHAM, Multi-HMR, W-HMR, and PLIKS (Natraj et al., 2026). Scoring evaluated nine lower-body joints (pelvis, bilateral hips, knees, ankles, and toes), with PA-MPJPE computed per valid frame.

Method	High resolution	Low resolution	Overall (valid frames)
CameraHMR	70.6 ± 15.2 mm	97.9 ± 25.5 mm	83.0 ± 24.6 mm (41,457)
WHAM	71.1 ± 13.8 mm	99.6 ± 22.0 mm	83.5 ± 22.7 mm (36,582)
W-HMR	70.8 ± 10.9 mm	99.3 ± 23.2 mm	83.7 ± 22.6 mm (34,088)
PLIKS	72.7 ± 14.2 mm	130.2 ± 54.5 mm	94.8 ± 45.1 mm (30,844)
Multi-HMR	74.6 ± 11.9 mm	112.2 ± 44.1 mm	81.7 ± 26.1 mm (18,801)

CameraHMR reconstructed the highest number of valid frames, with the second-lowest overall error, and degraded least at lower video resolutions. Although Multi-HMR achieved lower error on clean frames, it reconstructed fewer than half as many valid frames. Missing frames and joint errors both propagate into downstream kinematic calculations, so CameraHMR was selected for the released sequences.

scripts/camerahmr_pipeline.py assembles CameraHMR's per-frame outputs into the sequence files.

---

This repository

SCAI-SCI-Gait-Dataset/
├── .github/workflows/     CI: rejects participant data (data files, row-level tables)
├── assets/                figures and logos
├── metadata/
│   ├── README.md          clinical table, cohort counts, missing values
│   └── data_dictionary.csv  the release's dictionary: every column and file field
├── MoCap_Data/README.md   markers, force plates, EMG, trials.csv, gait parameters
├── SMPL_Data/README.md    SMPL keys, axes, timing, walking direction
├── SMPL_Videos/README.md  the renders
├── scripts/
│   ├── camerahmr_pipeline.py  CameraHMR per-frame outputs to SMPL sequences
│   └── README.md
├── visualize_mocap.py     viewer: markers and force plates
├── visualize_smpl.py      viewer: SMPL skeleton and mesh
├── requirements.txt
├── CITATION.cff
└── LICENSE.txt

The code that de-identified and assembled the release is kept in the SCAI Lab release toolkit, not here.

---

Installation

Python 3.10 or later:

git clone https://github.com/SCAI-Lab/SCAI-SCI-Gait-Dataset.git
cd SCAI-SCI-Gait-Dataset
pip install -r requirements.txt

This installs numpy, pandas, matplotlib and ezc3d (C3D reader). Re-running the CameraHMR pipeline also needs torch, pytorch3d and smplx; see scripts/README.md.

---

Getting the data

1. Get credentialed access on PhysioNet and sign the data use agreement.
2. Download the release and check it:

   cd RELEASE_v1.5_SCAI-SCI-Gait
   sha256sum -c manifest.sha256        # macOS: shasum -a 256 -c manifest.sha256

3. Keep the release outside this repository, and pass its paths to the viewers and scripts. This keeps patient data out of git. The examples below use DATA=/path/to/RELEASE_v1.5_SCAI-SCI-Gait.
4. Optional, SMPL mesh surfaces in the viewer: register at the SMPL project page (free for research). Put smpl_faces.npy, or the SMPL model files, in smpl_model/ in this repository.

---

Viewers

MoCap and force plates

python visualize_mocap.py $DATA/MoCap_Data/mocap<studyID>-<trial>.c3d
python visualize_mocap.py $DATA/MoCap_Data/mocap<studyID>-<trial>.c3d --save frame.png

Shows the 20 markers in 3D (pelvis green, left blue, right red), with Fx, Fy and Fz of both force plates below. A slider moves through the trial, and a dashed line on the force plots marks the frame shown. --save writes the middle frame to a PNG instead of opening a window.

SMPL skeleton and mesh

python visualize_smpl.py $DATA/SMPL_Data/v<studyID>_sagittal_<li|re>_defaced_smpl_sequence.pkl
python visualize_smpl.py $DATA/SMPL_Data/v<studyID>_sagittal_<li|re>_defaced_smpl_sequence.pkl --save frame.png

Shows the 24-joint skeleton (left blue, right red, midline green) next to the mesh. A slider moves through the frames. Without the SMPL faces the mesh is drawn as a point cloud.

Figures made with the viewers show real participant data. Share them only within the terms of the data use agreement.

---

Loading the data

MoCap (.c3d) with ezc3d

from pathlib import Path
import numpy as np
from ezc3d import c3d

DATA = Path("/path/to/RELEASE_v1.5_SCAI-SCI-Gait")
trial = sorted((DATA / "MoCap_Data").glob("*.c3d"))[0]

c = c3d(str(trial))
P = c["parameters"]

labels = P["POINT"]["LABELS"]["value"]            # the 20 protocol markers
xyz = c["data"]["points"][:3]                     # (3, 20, n_frames), mm, Z up, NaN = missing
rate = float(P["POINT"]["RATE"]["value"][0])      # 100, 200 or 246 Hz

# vertical force on plate 1, calibrated
fp = P["FORCE_PLATFORM"]
ch = fp["CHANNEL"]["value"][:, 0] - 1             # 0-based analog channels of plate 1
raw = c["data"]["analogs"][0, ch, :]              # volts
fz = (fp["CAL_MATRIX"]["value"][:, :, 0] @ raw)[2]  # N, force on the plate: negative under load

print(f"{len(labels)} markers, {xyz.shape[2]} frames at {rate:g} Hz")

SMPL sequences (.pkl)

import pickle
import numpy as np

path = sorted((DATA / "SMPL_Data").glob("*.pkl"))[0]
with open(path, "rb") as f:
    seq = pickle.load(f)                          # dict of NumPy arrays

t = seq["frame_ids"] / 15.0                       # seconds; frames may be skipped
joints = seq["joints"]                            # (F, 24, 3), m, Y up, root-relative
pose = np.hstack([seq["global_orient"], seq["body_pose"]])   # (F, 72) axis-angle
betas = seq["betas"].mean(axis=0)                 # shape varies per frame: average it

print(f"{len(t)} frames over {t[-1] - t[0]:.1f} s")

The walking direction of each sequence is in trials.csv (direction), not in the file.

Tables

import pandas as pd

meta = pd.read_csv(DATA / "metadata/medical_metadata.csv")
trials = pd.read_csv(DATA / "trials.csv")
gait = pd.read_csv(DATA / "gait_parameters.csv")

# walking speed by neurologic category
g = gait[gait["status"] == "ok"].merge(meta, on="studyID")
print(g.groupby("neurologic_category")["walking_speed_m_s"].describe())

---

Known limitations

• No trial-level link between video and MoCap. The release does not record which MoCap trial each SMPL sequence and render corresponds to. They share the same participant and recording session.
• No gait events in the C3D files. The EVENT group holds only acquisition triggers. gait_parameters.csv detects strides directly from marker trajectories.
• EMG is raw (not filtered or rectified). In 137 trials the analog channels carry generic names (Channel_NN). trials.csv gives the force-plate channel map and the monitored EMG muscles of every trial.
• Marker gaps. 37 trials have all 20 markers in fewer than 90% of frames (markers_complete_pct).
• One participant has motion capture but no SMPL sequence, and one SMPL sequence has no render.
• Empty cells in the clinical table indicate the value was not recorded.
• Single-center cohort. The cohort was recruited from a single specialized national rehabilitation center, and the majority of injuries are incomplete (AIS D). Characteristics may differ across other clinical populations.

---

Citation

Please cite the paper that presents the markerless biomechanics pipeline and the mesh recovery benchmark on this dataset:

@article{natraj2026monocular,
  title         = {Monocular markerless biomechanics for clinically interpretable gait assessment in spinal cord injury},
  author        = {Natraj, Shreyasvi and Ruepp, Mathieu and Li, Yanke and Riener, Robert and Eriks-Hoogland, Inge and Paez-Granados, Diego},
  journal       = {arXiv preprint arXiv:2610.05552},
  year          = {2026},
  eprint        = {2610.05552},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2610.05552},
  url           = {https://arxiv.org/abs/2610.05552}
}

and the earlier study of this cohort, which assessed monocular 3D pose estimation of gait in SCI:

@article{natraj2025sci3dpose,
  title     = {3D pose estimation for scalable remote gait kinematics assessment},
  author    = {Natraj, Shreyasvi and Messmer, T{\'e}mi and Fujii, Yoshiori and Suzuki, Kenji and Riener, Robert and Eriks-Hoogland, Inge and Paez-Granados, Diego},
  journal   = {npj Digital Medicine},
  year      = {2025},
  publisher = {Nature Publishing Group UK},
  doi       = {10.1038/s41746-025-02211-y},
  url       = {https://www.nature.com/articles/s41746-025-02211-y}
}

If you use the dataset or the code, please cite the dataset:

@misc{scai_sci_gait_dataset,
  title        = {{SCAI-SCI Gait Dataset}: laboratory motion capture, force plates, EMG and video-derived SMPL body models of walking in people with spinal cord injury},
  author       = {Paez-Granados, Diego and Natraj, Shreyasvi and Kupper, Lucas and Ruepp, Mathieu Vincent and Amrein, Sabrina and Eriks-Hoogland, Inge},
  howpublished = {PhysioNet},
  year         = {2026},
  note         = {Version 1.0},
  doi          = {10.13026/TODO},
  url          = {https://physionet.org/content/scai-sci-gait/1.0.0/}
}

PhysioNet also asks you to cite the platform: Goldberger A, Amaral L, Glass L, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 101(23):e215–e220 (2000).

CITATION.cff holds the dataset citation and both papers.

---

License

The code in this repository is licensed under the terms in LICENSE.txt.

The data are available exclusively for non-commercial scientific research, under the PhysioNet credentialed data use agreement. It strictly forbids commercial use, redistribution outside PhysioNet, and any attempt to identify or re-identify a participant.

Process the data locally. Do not upload it to services that share uploads automatically (for example, the hosted AddBiomechanics service shares every upload under CC BY 4.0). Using the SMPL sequences with an SMPL implementation also requires the SMPL model license. CameraHMR, which produced the sequences, is released for non-commercial research.

---

Acknowledgements

We thank the patients of the Swiss Paraplegic Centre who generously permitted the research use of their clinical data, and the gait laboratory clinical team who recorded and processed the assessments: Benjamin Hirsch, Kathrin Hartmann, and Eleonora Bronckers.

We thank the Clinical Trial Unit of the Swiss Paraplegic Centre (CTU SPC) for data management and quality monitoring, and Pascal Schlegel for data revision for publication.

---

Funding

This study was supported and partially funded by the Schweizer Paraplegiker-Stiftung (Swiss Paraplegic Foundation) and the ETH Zürich Foundation (Project 2021-HS-348), within the Digital Transformation in Personalized Healthcare initiative for individuals with spinal cord injury.

---

Authors and contributions

Author	Affiliation	Contributions (CRediT)
Diego Paez-Granados (Principal Investigator)	Spinal Cord Artificial Intelligence (SCAI) Lab, ETH Zürich, Zürich, Switzerland; Swiss Paraplegic Research, Swiss Paraplegic Centre, Nottwil, Switzerland; University of Tsukuba, Tsukuba, Ibaraki, Japan	Conceptualization, Methodology, Funding acquisition, Supervision, Project administration, Writing – review & editing
Shreyasvi Natraj (Lead Author & Software Lead)	Spinal Cord Artificial Intelligence (SCAI) Lab, ETH Zürich, Zürich, Switzerland; Swiss Paraplegic Research, Swiss Paraplegic Centre, Nottwil, Switzerland	Data curation, Formal analysis, Methodology, Software, Visualization, Writing – original draft
Lucas Kupper	ETH Zürich, Zürich, Switzerland	Data curation, Formal analysis
Mathieu Ruepp	Spinal Cord Artificial Intelligence (SCAI) Lab, ETH Zürich, Zürich, Switzerland; Swiss Paraplegic Research, Swiss Paraplegic Centre, Nottwil, Switzerland	Data curation, Formal analysis
Sabrina Amrein	ETH Zürich, Zürich, Switzerland	Data curation, Formal analysis
Inge Eriks-Hoogland (Clinical Lead & Study Sponsor)	Swiss Paraplegic Research, Swiss Paraplegic Centre, Nottwil, Switzerland; Faculty of Health Sciences and Medicine, University of Lucerne, Lucerne, Switzerland; Swiss Paraplegic Centre, Nottwil, Switzerland	Conceptualization, Investigation, Resources, Supervision

Lucas Kupper, Mathieu Ruepp, and Sabrina Amrein extracted the clinical and biomechanical data and conducted preliminary exploration and analysis. Inge Eriks-Hoogland is the clinical study sponsor and clinical project lead. Diego Paez-Granados is the Principal Investigator. Shreyasvi Natraj is the lead author and developer of the software and processing pipelines.

---

Contact

For questions regarding data access, credentialing on PhysioNet, technical issues, or the codebase:

• Repository Issues: Please open an issue in this repository for code questions or bug reports.
• Corresponding Authors:
  • Shreyasvi Natraj (Lead Author & Software Lead): shreyasvi.natraj@hest.ethz.ch
  • Diego Paez-Granados (Principal Investigator): diego.paez@hest.ethz.ch