DirtyMoCap: Robust Motion Capture from Unconstrained Markers
Paper • 2609.19927 • Published
HKMALA-Motion research access
Use your own Hugging Face account with a verified primary email address. Read the complete license below, fill in the registration fields and explicitly accept the terms. Access requests are approved automatically.
The complete license governing your request is shown below. The registration fields and your Hugging Face username and primary email are used for access and license administration and rights protection, and retained only as necessary for those purposes. Use only your own account.
License text SHA256: e3b0f033ad568e7353a3802b9fb11dd1007b2cb0d123ce5c42bbbf8f3dfdac0e
HKMALA Motion Noncommercial Research Data License
Licensor: Westlake University, acting through its Endless AI Lab. Dataset: HKMALA-Motion. This license governs each registered individual’s access to the released data for noncommercial computational research on human motion and directly related teaching. Redistribution and commercial use require separate permission.
1 Acceptance and scope
1.1 Access requires personal registration, verification of your registered email address, display of these complete terms, and affirmative acceptance through an agreement box that is unchecked by default. Download access is provided after these steps without manual application review. Acceptance grants a license to you personally; no institutional or paper signature is required. Your acceptance does not represent your employer’s or institution’s signature or agreement. Other users must register separately.
1.2 The Dataset comprises only files identified in each batch’s release manifest: numerical marker trajectories provided as point cloud files; approved source-provided martial-arts style and routine labels and technical metadata provided as Excel files; fitted SMPL-H body and hand poses, global orientation, translation and shape parameters provided as binary files; and dataset documentation provided as documents. Each dataset version is identified by its actual public release date in YYYY-MM-DD format. Direct personal identifiers are removed from released trajectory files and metadata, except for performer names whose public disclosure has been expressly consented to by the relevant performer. Original source filenames remain excluded. Binary files contain fitted parameter data, not model weights or human-body model assets. The Dataset excludes third-party model assets, algorithm code, model weights, original videos, photographs, audio, and complete original FBX files. Each manifest records the date-based dataset version and included file categories. Required source-data, participant and third-party authorizations must be verified before each release; shape parameters additionally require a privacy review before inclusion.
2 Research permission and commercial use
2.1 The Licensor grants you a free, limited, nonexclusive, nontransferable, nonsublicensable license solely for noncommercial computational research on human motion, including algorithm development, training, testing and evaluation, and teaching directly related to those purposes. A grant or research funding does not by itself make a use commercial. Eligibility depends on the actual purpose, rather than whether you belong to a university, company, or nonprofit organization.
2.2 Commercial use is prohibited, including internal R&D, evaluation or training directed toward commercial products, paid services, or customer engagements, and commercial exploitation of a model trained using the Dataset. This restriction does not assign ownership in the trained model. Additional commercial permission requires a separate written agreement and can cover only rights available under the IGA agreement and third-party licenses; it is not an automatic license upgrade.
3 Registration and accounts
3.1 Provide your accurate full name, current institution or independent-researcher status, email address and intended research use, and update material changes. Verify your registered email address. Eligibility depends on your actual use under Section 2, without manual application review. Keep your account, credentials, download tokens and restricted links private; they may not be shared or used by another person. The Licensor shall retain your registration details, acceptance date and time, the complete license text you accepted, and the dataset versions and batches you download. The accepted text shall be available to you for retention. These records are used for access and license administration and rights protection, and retained only as necessary for those purposes.
4 Internal processing and storage
4.1 You may clean, crop, normalize, transform or convert the Dataset internally for permitted research, and store working copies on multiple devices you control. Use reasonable access controls. A secured cloud processor may mechanically process data solely on your behalf, under confidentiality and deletion controls, without independent use or access by collaborators. You remain responsible for authorized processing and must ensure that the provider does not retain data beyond your permitted use.
5 No redistribution
5.1 Except for processing expressly permitted by Section 4, do not distribute, transfer, sublicense, sell, upload, mirror or otherwise provide the Dataset, subsets, or transformed versions from which source motions remain recoverable. Do not publish underlying numerical motion records, parameter files or downloadable 3D animation files. Research visualizations are permitted only as described in Section 7.1. Collaborators must obtain their own access from the official source, even if all have accepted this license; their acceptance does not authorize transfers between users. Publishing a new format, altered sampling rate or different body representation does not remove this prohibition where the underlying motions remain recoverable.
6 Third-party rights
6.1 This license covers only data and additional content that the Licensor has authority to license, and does not grant rights to use SMPL, SMPL-H or other third-party models. Users requiring those models must obtain the corresponding authorization themselves.
6.2 No SMPL, SMPL-H, MANO or other third-party model assets are transferred under this license. Obtain required assets from authorized sources and comply with each applicable model license. This license neither enlarges third-party permissions nor certifies compatibility with every third-party license or intended use.
7 Publication attribution and participants
7.1 Subject to participant and third-party rights, you may publish research statistics, graphs, still images and reasonable short rendered videos, including downloadable supplementary videos. They must not contain numerical motion records, parameter files or 3D animation files, or serve as a substitute dataset from which the released motion records can be reconstructed. Credit HKMALA-Motion and Westlake University / Endless AI Lab; acknowledge International Guoshu Association Limited (IGA) and the Institute of Chinese Martial Studies Limited (ICMS) for facilitating access to the source recordings. Cite DirtyMoCap: Robust Motion Capture from Unconstrained Markers (2026, arXiv:2609.19927) and applicable SMPL-H references and notices. Do not imply endorsement.
7.2 Do not attempt to reidentify participants whose identities have not been expressly disclosed by the Licensor with their consent. Consented disclosure of a name does not authorise biometric recognition, surveillance, misleading impersonation or unlawful privacy intrusion. Use of de-identified sequence codes does not guarantee full anonymization; shape parameters and motion patterns may retain identifying information and must be protected accordingly. The Dataset contains intermediate production data for computational research. Fitted reconstructions are algorithm outputs without manual post-refinement or final archival approval: they are research estimates, not verified ground truth or authoritative martial-arts exemplars. Records with clear data errors may be excluded from a release. For formally curated archival materials, contact IGA and ICMS.
8 Duration and breach
8.1 Download-link expiry does not terminate a compliant license. New terms apply only when you subsequently accept them, without retroactively changing the terms you previously accepted. An uncured material breach permits termination after notice and 30 days to cure. Immediate suspension may address a concrete urgent privacy, security or third-party-rights risk, with reasons communicated promptly. You may provide an explanation or evidence of remediation and request review of a suspension or proposed termination.
8.2 On termination, stop affected uses and delete affected copies, including processor copies. Retain only legally required archives, with restricted access and no further research use. Notify the Licensor when deletion is complete. Lawfully published research statistics and visualizations need not be withdrawn solely because this license terminates; this does not excuse correction or removal required to address an actual rights violation or binding legal order.
9 Warranties liability and governing law
9.1 The Dataset is provided as is, without promised accuracy, fitness, uninterrupted availability or freedom from third-party claims, to the extent permitted by law. Liability for indirect or consequential loss is excluded to that extent. These limits do not exclude liability that applicable law prohibits excluding, including fraud or other nonexcludable responsibility.
9.2 Hong Kong law governs this license, with exclusive jurisdiction of Hong Kong courts. Its English text prevails over the Chinese translation if they conflict. Any invalid provision is limited only as necessary; the remaining provisions continue to apply.
Website Acceptance Text
I have read and agree to the HKMALA Motion Noncommercial Research Data License and confirm that my declared use complies with it.
Log in or Sign Up to review the conditions and access this dataset content.
HKMALA-Motion
HKMALA-Motion provides numerical martial-arts motion data and DirtyMoCap-fitted SMPL-H motion parameters for noncommercial computational research on human motion and directly related teaching.
The source motion data were obtained from International Guoshu Association Limited (IGA) with authorization. We acknowledge IGA and the Institute of Chinese Martial Studies Limited (ICMS) for facilitating access to the source recordings.
Access
Sign in with your own Hugging Face account using a verified primary email address. Read the complete license displayed in the access form, provide your accurate full name, current institution or independent-researcher status and intended research use, then explicitly check the agreement box. Access requests are approved automatically; no institutional signature or manual application review is required. Access is personal, and accounts, tokens and download links must not be shared.
The registration fields, Hugging Face username and primary email are used for access and license administration and rights protection. They are retained only as necessary for those purposes, as described in Section 3.1 of the license.
Data and releases
The first public release contains 124 recordings, 1,238,717 frames, approximately 2.90 hours, and 1.90 GiB of compressed sequence data. Its release date and dataset version are 2026-11-02. manifest.json identifies the files, categories, version and checksums.
Martial-arts style and routine labels are preserved as provided. Ten recordings have no source-provided routine label; their routine entries are blank in metadata.xlsx and null in manifest.json. Performer names and original source filenames are excluded.
See DATA_FORMAT.md for array definitions, coordinate and SMPL-H conventions, and reading/reconstruction examples. The data use meters and Y up, with each recording's original nominal FPS. No resampling or manual correction has been applied.
The provided data categories are numerical marker trajectories, fitted SMPL-H pose/orientation/translation/shape parameters, approved martial-arts style and routine labels, technical metadata and documentation. Performer names and original source filenames are excluded from the released data. Body-model assets, model weights, original audiovisual recordings and complete original FBX files are excluded.
Fitted reconstructions are algorithm-generated estimates and have not been manually corrected. They do not represent or assess the performers' actual martial-arts skill.
A small number of frames may exhibit localized body or limb twisting, pose discontinuities or inaccurate hand poses. We are actively improving the reconstruction algorithm to reduce these artifacts.
They are not verified ground truth or authoritative martial-arts exemplars, and reconstruction artifacts must not be interpreted as evidence of a performer's mistakes or limitations. For formally curated archival materials, contact IGA and ICMS.
Download and verify
The repository is public with gated downloads. Complete the personal registration fields and accept the dataset license in your browser; access is approved automatically. Authenticate the same approved account to download:
hf auth login
hf download aLongZJU/HKMALA-Motion --repo-type dataset --local-dir assets/hkmala_motion
cd assets/hkmala_motion
sha256sum --check SHA256SUMS
Complete the personal access form and license agreement before your first download, then authenticate the same account for command-line downloads. Do not share your account, token or downloaded records.
License and attribution
The dataset uses the custom HKMALA Motion Noncommercial Research Data License. Read the complete terms below before requesting access. Commercial use and redistribution require separate permission. Body models and other third-party resources require their own authorizations.
Credit HKMALA-Motion and Westlake University / Endless AI Lab; acknowledge International Guoshu Association Limited (IGA) and the Institute of Chinese Martial Studies Limited (ICMS). Cite DirtyMoCap: Robust Motion Capture from Unconstrained Markers (2026, arXiv:2609.19927) and applicable SMPL-H references and notices as required by the license.
Complete license
HKMALA Motion Noncommercial Research Data License
Licensor: Westlake University, acting through its Endless AI Lab. Dataset: HKMALA-Motion. This license governs each registered individual’s access to the released data for noncommercial computational research on human motion and directly related teaching. Redistribution and commercial use require separate permission.
1 Acceptance and scope
1.1 Access requires personal registration, verification of your registered email address, display of these complete terms, and affirmative acceptance through an agreement box that is unchecked by default. Download access is provided after these steps without manual application review. Acceptance grants a license to you personally; no institutional or paper signature is required. Your acceptance does not represent your employer’s or institution’s signature or agreement. Other users must register separately.
1.2 The Dataset comprises only files identified in each batch’s release manifest: numerical marker trajectories provided as point cloud files; approved source-provided martial-arts style and routine labels and technical metadata provided as Excel files; fitted SMPL-H body and hand poses, global orientation, translation and shape parameters provided as binary files; and dataset documentation provided as documents. Each dataset version is identified by its actual public release date in YYYY-MM-DD format. Direct personal identifiers are removed from released trajectory files and metadata, except for performer names whose public disclosure has been expressly consented to by the relevant performer. Original source filenames remain excluded. Binary files contain fitted parameter data, not model weights or human-body model assets. The Dataset excludes third-party model assets, algorithm code, model weights, original videos, photographs, audio, and complete original FBX files. Each manifest records the date-based dataset version and included file categories. Required source-data, participant and third-party authorizations must be verified before each release; shape parameters additionally require a privacy review before inclusion.
2 Research permission and commercial use
2.1 The Licensor grants you a free, limited, nonexclusive, nontransferable, nonsublicensable license solely for noncommercial computational research on human motion, including algorithm development, training, testing and evaluation, and teaching directly related to those purposes. A grant or research funding does not by itself make a use commercial. Eligibility depends on the actual purpose, rather than whether you belong to a university, company, or nonprofit organization.
2.2 Commercial use is prohibited, including internal R&D, evaluation or training directed toward commercial products, paid services, or customer engagements, and commercial exploitation of a model trained using the Dataset. This restriction does not assign ownership in the trained model. Additional commercial permission requires a separate written agreement and can cover only rights available under the IGA agreement and third-party licenses; it is not an automatic license upgrade.
3 Registration and accounts
3.1 Provide your accurate full name, current institution or independent-researcher status, email address and intended research use, and update material changes. Verify your registered email address. Eligibility depends on your actual use under Section 2, without manual application review. Keep your account, credentials, download tokens and restricted links private; they may not be shared or used by another person. The Licensor shall retain your registration details, acceptance date and time, the complete license text you accepted, and the dataset versions and batches you download. The accepted text shall be available to you for retention. These records are used for access and license administration and rights protection, and retained only as necessary for those purposes.
4 Internal processing and storage
4.1 You may clean, crop, normalize, transform or convert the Dataset internally for permitted research, and store working copies on multiple devices you control. Use reasonable access controls. A secured cloud processor may mechanically process data solely on your behalf, under confidentiality and deletion controls, without independent use or access by collaborators. You remain responsible for authorized processing and must ensure that the provider does not retain data beyond your permitted use.
5 No redistribution
5.1 Except for processing expressly permitted by Section 4, do not distribute, transfer, sublicense, sell, upload, mirror or otherwise provide the Dataset, subsets, or transformed versions from which source motions remain recoverable. Do not publish underlying numerical motion records, parameter files or downloadable 3D animation files. Research visualizations are permitted only as described in Section 7.1. Collaborators must obtain their own access from the official source, even if all have accepted this license; their acceptance does not authorize transfers between users. Publishing a new format, altered sampling rate or different body representation does not remove this prohibition where the underlying motions remain recoverable.
6 Third-party rights
6.1 This license covers only data and additional content that the Licensor has authority to license, and does not grant rights to use SMPL, SMPL-H or other third-party models. Users requiring those models must obtain the corresponding authorization themselves.
6.2 No SMPL, SMPL-H, MANO or other third-party model assets are transferred under this license. Obtain required assets from authorized sources and comply with each applicable model license. This license neither enlarges third-party permissions nor certifies compatibility with every third-party license or intended use.
7 Publication attribution and participants
7.1 Subject to participant and third-party rights, you may publish research statistics, graphs, still images and reasonable short rendered videos, including downloadable supplementary videos. They must not contain numerical motion records, parameter files or 3D animation files, or serve as a substitute dataset from which the released motion records can be reconstructed. Credit HKMALA-Motion and Westlake University / Endless AI Lab; acknowledge International Guoshu Association Limited (IGA) and the Institute of Chinese Martial Studies Limited (ICMS) for facilitating access to the source recordings. Cite DirtyMoCap: Robust Motion Capture from Unconstrained Markers (2026, arXiv:2609.19927) and applicable SMPL-H references and notices. Do not imply endorsement.
7.2 Do not attempt to reidentify participants whose identities have not been expressly disclosed by the Licensor with their consent. Consented disclosure of a name does not authorise biometric recognition, surveillance, misleading impersonation or unlawful privacy intrusion. Use of de-identified sequence codes does not guarantee full anonymization; shape parameters and motion patterns may retain identifying information and must be protected accordingly. The Dataset contains intermediate production data for computational research. Fitted reconstructions are algorithm outputs without manual post-refinement or final archival approval: they are research estimates, not verified ground truth or authoritative martial-arts exemplars. Records with clear data errors may be excluded from a release. For formally curated archival materials, contact IGA and ICMS.
8 Duration and breach
8.1 Download-link expiry does not terminate a compliant license. New terms apply only when you subsequently accept them, without retroactively changing the terms you previously accepted. An uncured material breach permits termination after notice and 30 days to cure. Immediate suspension may address a concrete urgent privacy, security or third-party-rights risk, with reasons communicated promptly. You may provide an explanation or evidence of remediation and request review of a suspension or proposed termination.
8.2 On termination, stop affected uses and delete affected copies, including processor copies. Retain only legally required archives, with restricted access and no further research use. Notify the Licensor when deletion is complete. Lawfully published research statistics and visualizations need not be withdrawn solely because this license terminates; this does not excuse correction or removal required to address an actual rights violation or binding legal order.
9 Warranties liability and governing law
9.1 The Dataset is provided as is, without promised accuracy, fitness, uninterrupted availability or freedom from third-party claims, to the extent permitted by law. Liability for indirect or consequential loss is excluded to that extent. These limits do not exclude liability that applicable law prohibits excluding, including fraud or other nonexcludable responsibility.
9.2 Hong Kong law governs this license, with exclusive jurisdiction of Hong Kong courts. Its English text prevails over the Chinese translation if they conflict. Any invalid provision is limited only as necessary; the remaining provisions continue to apply.
Website Acceptance Text
I have read and agree to the HKMALA Motion Noncommercial Research Data License and confirm that my declared use complies with it.
- Downloads last month
- 29