Hugging Face's logo

Datasets:
kanglabs
/
ub-MOJI

Tasks:
Video Classification
Modalities:
Tabular
Text
Video
Size:
1K - 10K
Tags:
video
fingerspelling
Japanese
sign-language
DOI:
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
29

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

The “ub-MOJI” dataset is provided by the AI Vision Laboratory at Tokyo Polytechnic University and is permitted for academic research purposes only.
By requesting access, you agree to the following conditions:

  • Commercial use is prohibited
  • Use for identification of individuals or invasion of privacy is prohibited
  • Redistribution, transfer, or sublicensing is prohibited
  • You are responsible for compliance with the laws of your country
  • You may be required to delete the dataset if you violate the terms

By submitting a request, you agree to the above and to the full Terms of Use.

本データセット「ub-MOJI」は、東京工芸大学映像情報処理研究室により提供されるもので、学術研究目的に限って使用が許可されます。利用にあたっては、以下の規約に同意していただく必要があります。

  • 営利目的での使用は禁止されています
  • 個人の識別やプライバシーの侵害を目的とした利用は禁止されています
  • 再配布・譲渡・サブライセンスは禁止されています
  • 利用機関は、該当国の法令にも従う責任を負います
  • 規約違反時には使用停止と削除が求められます

本データセットの利用を申請することで、上記および本規約全文の内容に同意したものとみなされます。

Log in or Sign Up to review the conditions and access this dataset content.

ub-MOJI

Overview

ub-MOJI is a Japanese fingerspelling video dataset designed to advance research in sign language recognition. The name "ub-MOJI" is inspired by the Japanese word for fingerspelling, yubimoji (指文字). The dataset consists of video recordings of fingerspelling gestures performed in Japanese Sign Language (JSL), systematically organized into three levels of linguistic granularity: single characters (individual kana), five-character sequences, and complete words.

This dataset aims to support research tasks such as fingerspelling recognition and temporal segmentation of fingerspelling sequences.

Please note that a portion of the dataset is not publicly available, as some participants did not provide consent for open release.

Download Instructions

We strongly recommend specifying a dataset version to ensure reproducibility. The version follows a date-based format like 25.05. See Versioning Policy for details.

Requirement

Before downloading the ub-MOJI dataset, you must agree to the Terms of Use and log in to your Hugging Face account:

uvx hf auth login

Using hf CLI

You can download a specific version to the ub-moji directory as follows:

uvx hf download kanglabs/ub-MOJI --repo-type dataset --local-dir ub-moji --revision {version}

Using Git

This method requires git-xet to be installed. Clone a specific version as follows:

git xet install
git clone https://huggingface.co/datasets/kanglabs/ub-MOJI -b {version} --depth 1

Using Python library

Install the library:

uv add datasets

Load the dataset:

from datasets import load_dataset

continuous_dataset = load_dataset("kanglabs/ub-MOJI", "continuous", revision="{version}")
isolated_dataset = load_dataset("kanglabs/ub-MOJI", "isolated", revision="{version}")

Data Structure

The ub-MOJI dataset is organized into three subsets, each corresponding to a different linguistic unit of Japanese fingerspelling:

  • isolated/videos/: individual kana characters (organized by subdirectories)
  • continuous/videos/sequences/: sequences of five kana characters (stored as flat files)
  • continuous/videos/words/: fingerspelled full words (stored as flat files)

Each sample is stored as an RGB video file in .mp4 format. For continuous/videos/sequences/ and continuous/videos/words/, corresponding .toml files provide frame-level temporal annotations. Supplementary metadata in .parquet format summarizes information for each subset (isolated/metadata.parquet, continuous/metadata.parquet).

File Naming Convention

Files under continuous/videos/sequences/ and continuous/videos/words/ follow the format: {content}_{participantID}_{yyyymm}_{take}.mp4

  • {content}: sequence of kana (e.g., aiueo) or a full word (e.g., kamakura)
  • {participantID}: participant identifier (e.g., 001)
  • {yyyymm}: recording year and month
  • {take}: take number (e.g., t001)

Most files under isolated/videos/ are generated from continuous/ recordings by cutting out the segment for each annotations.toml entry, and are named after that entry's annotation_id instead: {annotation_id}.mp4 (e.g., ann_1a2b3c4d5e6f.mp4). Because these files carry no participant/date/take information in their name, that lineage is tracked via the annotation_id and source_video columns of isolated/metadata.parquet (see below) rather than the filename.

The only exception is the long class (the elongated-vowel mark, ー), which never appears in any continuous/ recording and therefore has no corresponding annotations.toml entry. Its 6 samples are standalone recordings that still use the legacy {content}_{participantID}_{yyyymm}_{take}.mp4 naming, with annotation_id/source_video left blank in the metadata.

Metadata and Annotation

  • isolated/metadata.parquet: sample-level metadata for isolated syllables
  • continuous/metadata.parquet: sample-level metadata for sequences and words
  • participants.csv: participant-level metadata (e.g., handedness, age group, etc.)
  • annotations.toml files provide time-series annotations for each character or word unit, facilitating temporal modeling tasks.

Data Fields

isolated/metadata.parquet

This file contains metadata for each isolated syllable sample. file_name is relative to the isolated/ directory. The columns are as follows:

Field Name Type Description
file_name str File path of the video sample, relative to isolated/
class_label str Fingerspelled syllable (e.g., "a")
participant_id int Participant identifier (e.g., 18)
recording_date str Year and month of recording (e.g., "202403")
fps float Frames per second (e.g., 30.0)
annotation_id str | null ID of the source annotations.toml segment this clip was cut from (e.g., "ann_1a2b3c4d5e6f"). null only for the long class, which has no continuous/ source (see File Naming Convention)
source_video str | null Source video in continuous/, relative to continuous/ (matches continuous/metadata.parquet's file_name). null under the same condition as annotation_id

continuous/metadata.parquet

This file contains metadata for each sequence/word sample. file_name is relative to the continuous/ directory. The columns are as follows:

Field Name Type Description
file_name str File path of the video sample, relative to continuous/
type str Linguistic unit type: sequence or word
classes List[str] Fingerspelled units (e.g., ["ka", "ma", "ku", "ra"])
participant_id int Participant identifier (e.g., 18)
recording_date str Year and month of recording (e.g., "202403")
fps float Frames per second (e.g., 30.0)

participants.csv

This file includes metadata about the participants involved in recording.

Field Name Type Description
participant_id int Participant identifier (e.g., "18")
age_group str Age decade group (e.g., "40" for age 40–49; "-1" if not provided)
gender int Gender category: 0=female, 1=male, "-1" if unspecified
dominant_hand int Dominant hand: 0=right, 1=left, "-1" if unspecified
experience_years str Years of sign language experience: one of "1-3", "4-6", ..., "51+" or "-1"
hearing_level int Self-reported hearing ability: 0 (no issue) to 4 (severe), or "-1"(unknown)
face_visibility int Face visibility consent: 1=agreed, 0=declined

hearing_level was collected by asking participants to choose the option that best describes their hearing:

Value Description
0 No difficulty hearing in daily life
1 Difficulty hearing soft sounds or conversation in noisy places; told that the TV volume is too loud
2 Occasional mishearing or difficulty following conversation
3 Can hear conversation only with a very loud voice or while wearing a hearing aid
4 Cannot hear conversation even while wearing a hearing aid

annotations.toml

This file contains time-aligned annotations for each fingerspelling video in the dataset. Each top-level TOML table represents a single video, identified by a unique video ID (e.g., "kamakura_018_202310_t001").

["<video_id>"]
duration = <float>
fps = <float>

[["<video_id>".annotations]]
annotation_id = "<str>"
label = "<str>"
label_id = <int>
segment = [<float>, <float>]
Field Name Type Description
"<video_id>" str Unique identifier for each video (includes participant and date metadata)
duration float Total duration of the video in seconds
fps float Frames per second (e.g., 60.0)
annotations List[dict] List of annotated segments for the video
annotation_id str Opaque unique identifier for the segment (e.g., "ann_1a2b3c4d5e6f"). Assigned once at creation and does not change when label/segment are later corrected.
label str Fingerspelled unit label (e.g., "ka", "ma")
label_id int Integer class index assigned to the label
segment List[float] Start and end time in seconds (e.g., [1.2, 2.8])

License and Terms of Use

The ub-MOJI dataset is available exclusively for non-commercial academic research.

Access to the dataset is gated on Hugging Face Datasets, and requires users to agree to the full terms of use before downloading.

By using the dataset, you agree to:

  • Use the data for non-commercial, academic purposes only
  • Not redistribute the data
  • Properly cite the dataset in any publications or derivative works

For the full license and conditions, please refer to License and Terms of Use.

Versioning Policy

The ub-MOJI dataset follows a date-based versioning scheme, formatted as YY.MM. For example, 25.05 refers to the May 2025 release.

Each release may include:

  • New samples (e.g., additional participants or word entries)
  • Annotation refinements
  • Structural or metadata schema changes

We recommend citing the specific version used in your experiments or publications to ensure reproducibility.

For details about changes in each release, please refer to the CHANGELOG.

Authors & Contributors

Authors

  • Tamon Kondo (Graduate School of Engineering, Tokyo Polytechnic University)
  • Ryota Murai (Graduate School of Engineering, Tokyo Polytechnic University)
  • Naoto Tsuta (Department of Engineering, Tokyo Polytechnic University)
  • Yousun Kang (Faculty of Engineering, Tokyo Polytechnic University)

Contributors

  • Renta Kuzuyama (Graduate School of Engineering, Tokyo Polytechnic University)
  • Natsuki Yamanaka (Faculty of Arts, Tokyo Polytechnic University)
  • Rei Aoki (Faculty of Arts, Tokyo Polytechnic University)
  • Fumitaka Ono (Faculty of Arts, Tokyo Polytechnic University)
  • Yonguk Lee (Faculty of Arts, Tokyo Polytechnic University)

Affiliations are listed as of the time the dataset was developed.

Acknowledgement

This dataset was made possible with the generous support of the following organizations and individuals:

Funding

Cooperation

We also express our sincere gratitude to all the participants who took part in the video recordings.

Citation

If you use the ub-MOJI dataset in your research, please cite the following reference(s):

@misc{ubmoji2025,
  title        = {ub-MOJI},
  author       = {Kondo, Tamon and Murai, Ryota and Tsuta, Naoto and Kang, Yousun},
  year         = {2025},
  url          = {https://huggingface.co/datasets/kanglabs/ub-MOJI},
  publisher    = {Hugging Face},
}

@InProceedings{Murai_2025_ICCV,
    author    = {Murai, Ryota and Tsuta, Naoto and Shin, Duk and Kang, Yousun},
    title     = {Point-Supervised Japanese Fingerspelling Localization via HR-Pro and Contrastive Learning},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2025},
    pages     = {4975-4982},
    doi       = {10.1109/ICCVW69036.2025.00516},
}

For a list of our publications that use the ub-MOJI dataset, please visit our project page.
Downloads last month
125
Total file size:
21.9 GB

Models trained or fine-tuned on kanglabs/ub-MOJI

Collection including kanglabs/ub-MOJI