WayenVan/how2sign-front-clips

How2Sign RGB Front Clips (VideoFolder)

This directory packages the frontal-view, sentence-level How2Sign RGB clips in Hugging Face VideoFolder format. It contains the official train, validation, and test splits with English sentence annotations.

Directory layout

how2sign_videofolder/
├── README.md
├── train/
│   ├── metadata.parquet
│   └── shard_001_032/ ... shard_032_032/
├── validation/
│   ├── metadata.parquet
│   └── shard_001_018/ ... shard_018_018/
└── test/
    ├── metadata.parquet
    └── shard_001_012/ ... shard_012_012/

The Parquet files contain only rows with an existing MP4 file:

Split	Videos / rows
train	31,047
validation	1,739
test	2,343

Metadata columns

• file_name: MP4 path relative to the split's metadata.parquet; this is the standard VideoFolder link field.
• mp4_path: portable MP4 path relative to the dataset root.
• split: train, validation, or test.
• VIDEO_ID, VIDEO_NAME: source-video identifiers.
• SENTENCE_ID, SENTENCE_NAME: sentence/clip identifiers.
• START, END: original temporal boundaries in the source video, in seconds.
• SENTENCE: English sentence annotation.

The MP4 files are already sentence-level clips. Do not apply START and END to them again; those timestamps refer to the original full-length videos.

Load locally

from datasets import load_dataset

dataset = load_dataset(
    "videofolder",
    data_dir="/path/to/how2sign_videofolder",
)

print(dataset)
print(dataset["train"][0])

For metadata-only processing without decoding videos:

import polars as pl

train = pl.read_parquet("how2sign_videofolder/train/metadata.parquet")
print(train.select("mp4_path", "SENTENCE").head())

Upload to Hugging Face Hub

Create a dataset repository and authenticate with hf auth login, then run:

hf upload YOUR_USERNAME/YOUR_DATASET \
  /path/to/how2sign_videofolder \
  . \
  --repo-type dataset

Annotation coverage

The source CSV files contain a few rows whose corresponding distributed clips are absent. These rows were excluded from metadata.parquet: 118 train rows, 2 validation rows, and 14 test rows. No distributed MP4 is missing an annotation.

License and citation

How2Sign is released for research use under the Creative Commons Attribution-NonCommercial 4.0 International license. See the official project page for the dataset terms.

@inproceedings{Duarte_CVPR2021,
  title     = {How2Sign: A Large-scale Multimodal Dataset for Continuous American Sign Language},
  author    = {Duarte, Amanda and Palaskar, Shruti and Ventura, Lucas and Ghadiyaram, Deepti and DeHaan, Kenneth and Metze, Florian and Torres, Jordi and Giro-i-Nieto, Xavier},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2021}
}