danielelvs/multilingual-islr-mediapipe

Multilingual ISLR MediaPipe Landmarks

Dataset Description

This dataset combines frame-level MediaPipe Holistic landmarks derived from four isolated sign language recognition (ISLR) resources: INCLUDE-50, KSL, MINDS-Libras, and LIBRAS-UFOP. It provides a common tabular schema for research on landmark selection, temporal modeling, signer-independent evaluation, and multilingual transfer learning.

The release contains landmarks rather than source RGB videos. Every frame has 543 landmarks with three coordinates each: 468 face landmarks, 21 landmarks for each hand, and 33 pose landmarks. Labels, sample provenance, and repeated frame metadata are normalized into separate tables.

This is a harmonized collection of derived artifacts. It does not assert that the source corpora share recording conditions, coordinate reference systems, class semantics, participant populations, or licenses.

• Languages: Brazilian Sign Language (Libras), Indian Sign Language (ISL), and Korean Sign Language (KSL)
• Task: isolated sign language recognition from landmark sequences
• Unit of observation: one video frame
• Landmark extractor: MediaPipe Holistic
• Coordinates per frame: 1,629 (543 landmarks × 3 axes)
• Total columns in frames.csv: 1,641
• Processing repository: danielelvs/islr-subset
• Hugging Face collection: ISLR Subsets Datasets

Release Statistics

The definitive number of frames, samples, classes, and signers in this release is recorded under summary and totals in manifest.json, generated from the CSVs actually used to build the publication. This avoids reproducing statistics from an older extraction or a different train/validation/test split.

The publication script checks the same experimental coverage described in the dissertation before exporting a complete release:

Dataset	Expected samples	Expected signer IDs / evaluation split
INCLUDE-50	958	train 689; validation 77; test 192
KSL	1,229	20 signer IDs
MINDS-Libras	1,158	12 signer IDs
LIBRAS-UFOP	3,040 segments	5 signer IDs

These are experimental coverage targets, not a substitute for inspecting manifest.json and samples.csv in the published artifact. Raw source video counts, frame counts, and availability depend on the actual extraction.

INCLUDE-50 split

Only INCLUDE-50 includes explicit train, val, and test split labels in this bundle. They are copied from the verified input CSV, not generated randomly. Consult manifest.json and group samples.csv with frames.csv to verify the published counts. Never split frames of a video between partitions.

KSL, MINDS-Libras, and LIBRAS-UFOP have an empty split field: construct their evaluation partitions using signer and sequence groups rather than splitting frames randomly.

Dataset Structure

frames.csv             # one row per frame
classes.csv            # class labels and provenance
samples.csv            # video/segment provenance
data_dictionary.csv    # definition of every published column
manifest.json          # inputs, counts, and export status
README.md

The frame key is (sample_id, frame_id). Read identifiers as strings to preserve leading zeros. Join frames.csv to the lookup tables using class_id and sample_id, never labels, filenames, or source-local IDs.

frames.csv

The first 12 columns are canonical metadata:

Column	Type	Nullable	Description
dataset	string	no	include50, ksl, minds, or ufop.
class_id	string	no	Global class key; references classes.csv.
sample_id	string	no	Global video/segment key; references samples.csv.
sequence_id	string	no	Groups segments from one original sequence; equals sample_id outside UFOP.
signer_id	string	yes	Source-prefixed signer ID; unavailable for INCLUDE-50.
frame_id	integer	no	Non-negative frame index within the sample; gaps are preserved.
source_frame_id	integer	yes	Original UFOP frame index; empty elsewhere.
split	string	yes	train, val, or test for INCLUDE-50; empty elsewhere.
missing_hand_0	boolean	no	True if any coordinate from hand 0 is missing.
missing_hand_1	boolean	no	True if any coordinate from hand 1 is missing.
missing_pose	boolean	no	True if any pose coordinate is missing.
missing_face	boolean	no	True if any face coordinate is missing.

The remaining 1,629 columns follow this order:

face_0_x ... face_467_z
hand_0_0_x ... hand_0_20_z
pose_0_x ... pose_32_z
hand_1_0_x ... hand_1_20_z

Every landmark has _x, _y, and _z coordinates. Empty fields represent missing values. Zero is a valid coordinate.

classes.csv

Column	Description
class_id	Global key in the form dataset::source_id.
source_class_id	Class identifier in the extracted source; INCLUDE-50 IDs 0–49 are assigned by the extraction notebook.
label	Official benchmark name for INCLUDE-50; original sign text for KSL; name parsed from the source video filename for MINDS. Empty for UFOP without a verified mapping.
category_name	Original INCLUDE-50 thematic category.
source_category_id	Original UFOP category code.
source_class_key	INCLUDE-50 category/raw-name key generated in extraction; UFOP source sign key.
source_class_id_in_category	Original UFOP within-category sign ID.

Class label provenance

• INCLUDE-50 label uses official_sign, not the possibly different raw-folder sign. For example, the benchmark uses Shop, Large, and Small where source folders use Store or Shop, big large, and small little. The original folder-derived key remains in source_class_key.
• MINDS-Libras names are parsed from source video filenames and checked against their class ID and signer ID.
• KSL keeps source-provided sign names (including original spellings); no automatic spelling corrections or translations are applied.
• LIBRAS-UFOP label remains empty until a validated ID-to-label inventory is available.

These class names are source-local: matching words across sign languages do not imply equivalent signs.

samples.csv

Column	Description
sample_id	Global video or segment key.
source_sample_id	Original sample ID when supplied.
source_sequence_id	Original sequence ID when supplied.
source_signer_id	Original signer ID for KSL, MINDS, and UFOP.
source_video_name	Original video name.
source_video_relpath	Original relative path for INCLUDE-50.
source_start_frame	Original inclusive UFOP segment start.
source_end_frame	Original inclusive UFOP segment end.

Global IDs contain a dataset prefix and percent-escaped source components separated by ::.

Data Processing

The publication pipeline harmonizes schemas without rounding, interpolation, normalization, or rescaling of finite coordinates. It:

1. validates each source header;
2. maps anatomical MINDS landmark names to MediaPipe indices;
3. creates global class, sample, sequence, and signer IDs;
4. moves repeated class and sample data into lookup tables;
5. converts empty and case-insensitive NaN fields to empty CSV values;
6. recomputes missing-landmark flags from the published coordinates; and
7. validates numeric values, keys, metadata consistency, and UFOP segment limits.

The ambiguous source field missing_hand is excluded. Derive missing_hand_0 OR missing_hand_1 or missing_hand_0 AND missing_hand_1 explicitly, depending on the required meaning.

Rebuild the release with notebooks/merge_mediapipe_datasets.ipynb. The generated manifest.json records counts, source-file digests when enabled, and processing notebook digest. SHA256SUMS protects the output artifact files. Metadata-only class-label repairs update their affected checksums without recomputing frames.csv.

Loading the Dataset

Each table is a Hugging Face configuration:

from datasets import load_dataset

repo_id = "danielelvs/multilingual-islr-mediapipe"
frames = load_dataset(repo_id, "frames", split="train")
classes = load_dataset(repo_id, "classes", split="train")
samples = load_dataset(repo_id, "samples", split="train")

Stream the large frame table when memory is limited:

frames = load_dataset(repo_id, "frames", split="train", streaming=True)
for row in frames.take(3):
    print(row["dataset"], row["sample_id"], row["frame_id"])

With pandas, read selected columns in chunks:

import pandas as pd

metadata = [
    "dataset", "class_id", "sample_id", "sequence_id", "signer_id",
    "frame_id", "source_frame_id", "split",
    "missing_hand_0", "missing_hand_1", "missing_pose", "missing_face",
]
for chunk in pd.read_csv("frames.csv", usecols=metadata, chunksize=50_000):
    print(chunk.groupby("dataset").size())

For relational joins:

frames = pd.read_csv("frames.csv", dtype={"class_id": "string", "sample_id": "string"})
classes = pd.read_csv("classes.csv", dtype="string")
samples = pd.read_csv("samples.csv", dtype="string")

data = (
    frames
    .merge(classes, on="class_id", how="left", validate="many_to_one")
    .merge(samples, on="sample_id", how="left", validate="many_to_one")
)

Recommended Evaluation

Frames from the same sample are temporally related and must stay in one partition. UFOP segments sharing a sequence_id must also remain together.

• Use the provided split for INCLUDE-50.
• Use signer-independent grouped evaluation for KSL, MINDS-Libras, and LIBRAS-UFOP, such as leave-one-person-out validation.
• Create splits independently per source unless the research question defines a cross-dataset protocol.
• Report results per source because pooled scores can conceal differences in language, label space, participants, and capture conditions.

This release does not provide a verified multilingual mapping between signs.

Limitations, Biases, and Privacy

• These are processed subsets available to this project and may omit items from the original corpora.
• MediaPipe missingness can correlate with motion, occlusion, camera angle, skin appearance, clothing, background, and capture quality, creating systematic bias.
• Source acquisition and extraction settings can differ. Identical column names do not establish measurement equivalence.
• Coordinates are not calibrated 3D world positions unless the relevant source processing record explicitly establishes that property.
• Labels are source-provided and have not been linguistically normalized, translated, or validated across languages.
• Participant coverage is limited; results may not generalize to unseen signers, dialects, signing styles, cameras, or environments.
• INCLUDE-50 has no reliable signer ID in the available table, so signer-independent evaluation cannot be reconstructed here.
• Face landmarks and body motion can retain biometric and behavioral information. Removing RGB pixels does not make the data anonymous.

Suitable research uses include isolated sign recognition, temporal modeling, landmark subset selection, missing-data robustness, and domain-shift analysis. Do not use the release for biometric identification, signer re-identification, surveillance, high-stakes decisions about people, or claims of linguistic equivalence across sign languages. An isolated-sign classifier is not a complete sign language translation system.

License and Source Terms

This harmonized landmark release is distributed under the MIT License by its maintainer. The release contains derived landmark coordinates rather than source RGB videos. Users must retain the source citations and comply with any attribution or use conditions that continue to apply to the original corpora.

Known source pages:

• INCLUDE repository
• KSL repository
• MINDS-Libras catalogue entry
• LIBRAS-UFOP thesis record

Citation

Cite the associated paper when using this release:

@article{dosSantos2025proper,
  title   = {Proper Body Landmark Subset Enables More Accurate and 5X Faster Recognition of Isolated Signs in LIBRAS},
  author  = {dos Santos, Daniele L. V. and Pereira, Thiago B. and Alves, Carlos Eduardo G. R. and Tello, Richard J. M. G. and Boldt, Francisco de A. and Paix{\~a}o, Thiago M.},
  journal = {arXiv preprint arXiv:2510.24887},
  year    = {2025},
  doi     = {10.48550/arXiv.2510.24887}
}

Cite every original source represented in the experiment as well. INCLUDE should at minimum be cited as:

@inproceedings{sridhar2020include,
  title     = {INCLUDE: A Large Scale Dataset for Indian Sign Language Recognition},
  author    = {Sridhar, Advaith and Ganesan, Rohith and Kumar, Pratyush and Khapra, Mitesh M.},
  booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},
  year      = {2020},
  doi       = {10.1145/3394171.3413528}
}

Use the source pages above to identify the exact KSL, MINDS-Libras, and LIBRAS-UFOP versions used in an experiment.

Reproducibility and Versioning

Each publication release should preserve the generated manifest.json, processing commit hash, MediaPipe version and settings per source, immutable input identifiers or checksums, exact source licenses and citations, and any filtering performed before extraction.

Counts in this card correspond to the current processed inputs. Regenerate the tables and statistics whenever inputs or harmonization code change.