STS-lexicon-contacts

STS-lexicon-contacts is a dataset of contact rich reconstructed Swedish Sign Language (STS) motion sequences, with semantically meaningful self-contacts identified from sign transcriptions. It contains 5 313 sequences and over 4 hours of SMPL-X motion data. The source sequences are taken from the Swedish Sign Language Lexicon (Svenskt teckenspråkslexikon). Each sequence includes SMPL-X parameters and accompanying lexicon and contact metadata.

Dataset Summary

Across the dataset, 7 516 contact pairs have been identified:

Contact category	Count
Hand-to-hand	1 834
Hand-to-body	2 577
Hand-to-face	3 105
Total	7 516

Contact pairs indicate which types of contact occur somewhere in a sequence, mapping a hand to a body region. They do not specify the exact contact frames. Each sequence also includes a signing interval: an estimated frame interval during which signing takes place, based on hand movement.

Data Structure

The dataset is distributed as a .zip archive. Each sequence is represented by two files with a shared name:

[sign_name]-[lexikon_id].sequence.npz
[sign_name]-[lexikon_id].json

Motion Sequence

The .sequence.npz file contains SMPL-X parameters as arrays. T denotes the number of frames in a sequence.

Field	Shape	Description
global_orient	(T, 3)	Global body orientation, in axis-angle form
body_pose	(T, 21, 3)	Body-joint poses, in axis-angle form
left_hand_pose	(T, 15, 3)	Left-hand joint poses, in axis-angle form
right_hand_pose	(T, 15, 3)	Right-hand joint poses, in axis-angle form
jaw_pose	(T, 3)	Jaw pose, in axis-angle form
leye_pose	(T, 3)	Left-eye pose, in axis-angle form
reye_pose	(T, 3)	Right-eye pose, in axis-angle form
expression	(T, 10)	Facial expression coefficients
transl	(T, 3)	Global body translation
betas	(10,)	Body-shape coefficients

Sequence Metadata

The corresponding .json file contains:

Field	Description
sign_name	Name of the sign
lexikon_id	Identifier in the lexicon
formbeskrivning	Natural language movement description in Swedish from the lexicon
transkription	Sign transcription from the lexicon
video_url	URL of the source lexicon video
same_form_ids	Lexicon IDs for signs with the same movement
sign_url	URL of the sign's lexicon entry
framerate	Sequence frame rate
sign_class	Thesis-specific sign class label, describes which hand is active
signing_interval	Estimated signing interval, with start and end frames, and frame count
contact_pairs	Contact-pair labels identified from the transcription

Reconstruction

The body meshes were reconstructed by combining holistic and hand-specific models:

• SMPLer-X was used for holistic human mesh recovery.
• WiLoR was used for hand reconstruction.

The body and hand reconstructions were combined at the wrist. The global rotations of the elbow and complete hand were used to derive the wrist's local pose, while the MANO finger poses were transferred directly to the SMPL-X body. Important to note that flat_hand_mean is True for these sequences.

Intended Uses

STS-lexicon-contacts is intended for research involving the Swedish Sign Language, body-mesh reconstruction, sign-language contact analysis, and contact-aware motion modelling and retargeting.

Limitations and Considerations

• The dataset contains only STS signs.
• Although the source material includes multiple signers, signer and dataset biases may still be present.
• The sequences are estimated reconstructions rather than ground-truth motion capture. Reconstruction geometry and motion may be faulty, including self-penetrations, missed contacts, and inaccurate poses.
• Hand-to-face contacts are particularly affected by depth ambiguity and frequently do not make full geometric contact with the face.
• Contact labels identify contact pair types at the sequence level, not precise contact timings. The contact labels may additionally contain false positive contacts, and exclude true positive contacts, due to ambiguities in lexical transcriptions.
• Signing intervals are estimates based on hand movement, and may include non-signing movement or exclude signing movement.

Users should account for these limitations when training models, evaluating contact detection, or drawing other conclusions.

License

This dataset, making use of source sequences from Svenskt teckenspråkslexikon, is published under CC BY-NC-SA 4.0.

Citation

When using this dataset, please reference both the accompanying thesis and the original data source, Svenskt teckenspråkslexikon:

@mastersthesis{sts_contacts_2026,
    title = {Contact {Estimation} for {Sign} {Language} {Motion} {Analysis}},
    school = {KTH, School of Electrical Engineering and Computer Science (EECS)},
    author = {Andersson, Nora},
    year = {2026},
}

@misc{svenskt_teckensprakslexikon_2023,
    address = {Stockholm},
    title = {Svenskt teckenspråkslexikon},
    url = {https://teckensprakslexikon.su.se/},
    language = {sv},
    urldate = {2026-06-15},
    publisher = {Stockholm University},
    year = {2023},
}

Acknowledgements

This dataset builds on the work and resources provided by:

• Svenskt teckenspråkslexikon, the source of the video data and lexical metadata.
• SMPLer-X, used for holistic human mesh recovery.
• WiLoR, used for hand reconstruction.