MySign

MySign: A High-Fidelity Motion-Capture Dataset for 3D Sign Generation in Bahasa Isyarat Malaysia (NeurIPS 2026, Evaluations and Datasets Track)

MySign is the first motion-capture dataset of isolated signs in Malaysian Sign Language (Bahasa Isyarat Malaysia, BIM). It contains 5,000 samples covering 1,000 glosses, each performed by five Deaf native signers, and is distributed as Filmbox (.fbx) skeletal animation retargeted to the SMPL-X body model.

Note: this dataset is not loadable via datasets.load_dataset(...) as a tabular split, because the data itself is binary .fbx. metadata.csv is an index over the FBX files; consumers must read the file at file_name to access the motion data.

Authors

Jiayu Shen^1,5, Kalin Stefanov², Lay-Ki Soon^1,5, Vee Yee Chong³, Alpha Agape Gopalai⁴, KokSheik Wong^1,5

¹ School of Information Technology, Monash University Malaysia ² Japan Advanced Institute of Science and Technology ³ Jeffrey Cheah School of Medicine & Health Sciences, Monash University Malaysia ⁴ School of Engineering, Monash University Malaysia ⁵ Action Lab Malaysia, Monash University Malaysia

Dataset Summary

• Modality: 3D skeletal motion capture, retargeted to SMPL-X (.fbx)
• Language: Malaysian Sign Language (BIM), ISO 639-3 xml; gloss labels are in English
• Signers: 5 Deaf native BIM signers (Signer001 … Signer005)
• Vocabulary: 1,000 glosses from the BIM Sign Bank, spanning 9 main categories and 46 subcategories
• Samples: 5,000 (5 signers × 1,000 glosses), fully balanced: every (signer, gloss) pair has exactly one .fbx file
• Capture: six-camera OptiTrack system at 200 Hz plus MANUS Prime 3 Data Gloves; ~15.57M synchronized frames
• Content: skeletal motion only; no video, audio, or facial texture
• License: CC BY-NC-SA 4.0
• Source repository: https://huggingface.co/datasets/mysigner/MySign

Repository Structure

MySign/
├── Signer001/
│   ├── Above.fbx
│   ├── Apologize.fbx
│   ├── Nine.001.fbx
│   └── ...
├── Signer002/
├── Signer003/
├── Signer004/
├── Signer005/
├── metadata.csv      # index over all .fbx files
├── croissant.json    # Croissant 1.0 + RAI 1.0 metadata (JSON-LD)
├── README.md
└── .gitattributes

metadata.csv Schema

metadata.csv is a UTF-8 CSV with a header row and 5,000 data rows. Each row corresponds to exactly one .fbx file in the repository.

Column	Type	Description
file_name	string	Repository-relative path to the .fbx, e.g. Signer001/Above.fbx.
gloss	string	Normalized gloss label (UPPERCASE). See Gloss Normalization.
signer_id	string	One of Signer001 … Signer005.
take	integer	Take index parsed from the file name: no suffix → 1, Blender-style duplicate suffix .001 → 2. Every (signer, gloss) pair has exactly one released file; 13 of the 5,000 files carry a .001 suffix from export naming and are the only released recording of their pair.

Example rows

file_name,gloss,signer_id,take
Signer001/Above.fbx,ABOVE,Signer001,1
Signer001/Apologize.fbx,APOLOGIZE,Signer001,1
Signer001/Nine.001.fbx,NINE,Signer001,2
"Signer002/Actor,_Actress.fbx",ACTOR / ACTRESS,Signer002,1
Signer002/1_hr.fbx,1 HOUR,Signer002,1
Signer004/Less(I).fbx,LESS(I),Signer004,1

Path quoting: some file_name values contain commas (e.g. Signer002/Actor,_Actress.fbx). The CSV is written with Python's default csv.writer, which quotes such fields automatically. Read it with any standard CSV parser (pandas, csv.DictReader).

Gloss Normalization

The gloss column is derived from each filename and normalized so that the same sign performed by different signers gets the same label. The normalization, applied by generate_metadata_remote.py, is:

1. Strip Blender suffix .001 / .002 (recorded into take).
2. Replace _ and - with spaces.
3. Treat , and ; as synonym separators, joined as WORD_A / WORD_B.
4. Uppercase.
5. Expand time-unit abbreviations: HR→HOUR, MIN→MINUTE, MTH→MONTH, WK→WEEK, YR→YEAR, SEC→SECOND (and plurals).
6. Normalize whitespace around parentheses: WORD (X) → WORD(X).
7. Strip unbalanced trailing ).
8. Plural→singular merge: when the corpus contains both forms of a word (e.g. COURSES and COURSE), the plural is rewritten to the singular. A small block-list keeps semantically distinct plurals (NEWS, SHORTS, MATHEMATICS) as plurals.
9. A small hand-curated override map (currently: COCHLEAR / IMPLANT → COCHLEAR IMPLANT) fixes cases that no general rule can fix safely.

A full audit of the plural→singular merges is printed by the script every time it runs.

.fbx Files

• Format: Autodesk Filmbox (binary .fbx), containing SMPL-X-aligned skeletal animation.
• One file per (signer, gloss) pair.
• The file_name column of metadata.csv is the canonical pointer to each file. The same path can be resolved as https://huggingface.co/datasets/mysigner/MySign/resolve/main/<file_name>.

Data Collection and Processing

Each signer was shown, sign by sign, a slide with the target gloss and the corresponding BIM Sign Bank reference video, and produced one take of the sign. The gloss label of every sample is therefore fixed at capture time against an authorized BIM Sign Bank entry. Processing steps:

1. Marker reconstruction in OptiTrack Motive 3.0.3.
2. SMPL-X retargeting with the Rokoko Retarget plugin v1.4.3 in Blender 4.5.2.
3. Anatomical range-of-motion clamping on the SMPL-X parameters to remove physically inadmissible joint configurations.

No data augmentation, mirroring, retiming, or motion editing was applied. Full details are in the paper and in the Responsible AI fields of croissant.json.

Ethics and Consent

All capture procedures were reviewed and approved by the Monash University Human Research Ethics Committee (MUHREC), Project ID 46717. Recruitment and consent materials were delivered primarily in BIM. Each participant gave written informed consent before their first session, participation was voluntary, and participants were compensated above the local minimum wage.

The released files contain no names, images, video, audio, or other direct identifiers; signers appear only as Signer001 … Signer005. Because re-identification from movement kinematics is possible in principle, the data should be treated as pseudonymous rather than fully anonymous.

Source Vocabulary

The 1,000-gloss vocabulary and the reference videos shown during capture come from the BIM Sign Bank, developed by the Malaysian Federation of the Deaf together with Guidewire Gives Back, and used under non-commercial research authorization. No BIM Sign Bank video, image, or audio is redistributed in MySign.

Croissant Metadata

croissant.json is a Croissant 1.0 (JSON-LD) description with Responsible AI (RAI 1.0) fields, validated with the NeurIPS Croissant checker. It declares:

• metadata.csv as a cr:FileObject of type text/csv, fetched directly from its resolve/main URL and verified by SHA-256 (no clone of the full repository is needed).
• A cr:RecordSet named signs exposing the four columns above.
• RAI metadata covering data collection, preprocessing, annotation, biases, limitations, intended use cases, social impact, and personal/sensitive information, plus provenance (prov:wasDerivedFrom, prov:wasGeneratedBy).

Limitation: Croissant has no native semantics for FBX. mlcroissant yields the index records only; consumers must use an FBX library (e.g. Autodesk FBX SDK, Blender) to read the motion data.

Usage

List records via metadata.csv

import pandas as pd

df = pd.read_csv(
    "https://huggingface.co/datasets/mysigner/MySign/resolve/main/metadata.csv"
)
print(df.head())
print(df["signer_id"].value_counts().sort_index())
print(df["gloss"].nunique(), "unique glosses")

Download a specific FBX

from huggingface_hub import hf_hub_download

local_path = hf_hub_download(
    repo_id="mysigner/MySign",
    filename="Signer001/Above.fbx",
    repo_type="dataset",
)
print(local_path)  # local cache path; open with your FBX library

Inspect via Croissant (mlcroissant)

import mlcroissant as mlc

ds = mlc.Dataset(
    "https://huggingface.co/datasets/mysigner/MySign/resolve/main/croissant.json"
)
print(ds.metadata.name, "—", ds.metadata.description[:80])

for i, rec in enumerate(ds.records(record_set="signs")):
    print(rec)
    if i >= 4:
        break

ds.records("signs") yields one Python dict per .fbx, with file_name, gloss, signer_id, take. The FBX bytes are not auto-loaded — read them yourself from file_name.

License

This dataset is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). You may share and adapt the work for non-commercial purposes, provided you give appropriate credit and distribute derivative works under the same license.

Citation

@inproceedings{shen2026mysign,
  title     = {MySign: A High-Fidelity Motion-Capture Dataset for 3D Sign Generation in Bahasa Isyarat Malaysia},
  author    = {Shen, Jiayu and Stefanov, Kalin and Soon, Lay-Ki and Chong, Vee Yee and Gopalai, Alpha Agape and Wong, KokSheik},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://huggingface.co/datasets/mysigner/MySign}
}

Limitations

• Isolated signs only: each FBX is one isolated, prompted sign. The dataset does not include continuous signing, sentences, classifier constructions, or non-manual markers (facial expression, mouthing, eye gaze).
• Five signers: although balanced across gender and four Malaysian ethnic backgrounds, five signers are too few to support strong claims about generalization across the full BIM-using population.
• Studio capture: all recordings come from a single motion-capture laboratory under controlled conditions.
• Processed data only: raw OptiTrack marker streams and raw glove streams are not released; SMPL-X retargeting and range-of-motion clamping may smooth some fine-grained signer-specific motion.
• Not auto-loadable: Hugging Face's Dataset Viewer cannot render .fbx, and datasets.load_dataset("mysigner/MySign") will not produce a usable split. Use the workflows above instead.
• Gloss labels are filename-derived: labels are normalized from filenames by a deterministic script; see Gloss Normalization.
• Not for high-stakes deployment: models trained on MySign should not be deployed as stand-alone interpreters in legal, medical, or educational settings without validation by the BIM-using community.