WLASL: A large-scale dataset for Word-Level American Sign Language (WACV 20' Best Paper Honourable Mention)

This repository contains the WLASL dataset described in "Word-level Deep Sign Language Recognition from Video: A New Large-scale Dataset and Methods Comparison".

Please visit the project homepage for news update.

Please star the repo to help with the visibility if you find it useful.

Download Original Videos

Download repo.

git clone https://github.com/dxli94/WLASL.git

Install youtube-dl for downloading YouTube videos.
Download raw videos.

cd start_kit
python video_downloader.py

Extract video samples from raw videos.

python preprocess.py

You should expect to see video samples under directory videos/.

Requesting Missing / Pre-processed Videos

Videos can dissapear over time due to expired urls, so you may find the downloaded videos incomplete. In this regard, we provide the following solution for you to have access to missing videos.

We also provide pre-processed videos for the full WLASL dataset on request, which saves troubles of video processing for you.

(a) Run

python find_missing.py

to generate text file missing.txt containing missing video IDs.

(b) Submit a video request by agreeing to terms of use at: https://docs.google.com/forms/d/e/1FAIpQLSc3yHyAranhpkC9ur_Z-Gu5gS5M0WnKtHV07Vo6eL6nZHzruw/viewform?usp=sf_link. You will get links to the missing videos within 72 hours.

File Description

The repository contains following files:

WLASL_vx.x.json: JSON file including all the data samples.
data_reader.py: Sample code for loading the dataset.
video_downloader.py: Sample code demonstrating how to download data samples.
C-UDA-1.0.pdf: the Computational Use of Data Agreement (C-UDA) agreement. You must read and agree with the terms before using the dataset.
README.md: this file.

Data Description

gloss: str, data file is structured/categorised based on sign gloss, or namely, labels.
bbox: [int], bounding box detected using YOLOv3 of (xmin, ymin, xmax, ymax) convention. Following OpenCV convention, (0, 0) is the up-left corner.
fps: int, frame rate (=25) used to decode the video as in the paper.
frame_start: int, the starting frame of the gloss in the video (decoding with FPS=25), indexed from 1.
frame_end: int, the ending frame of the gloss in the video (decoding with FPS=25). -1 indicates the gloss ends at the last frame of the video.
instance_id: int, id of the instance in the same class/gloss.
signer_id: int, id of the signer.
source: str, a string identifier for the source site.
split: str, indicates sample belongs to which subset.
url: str, used for video downloading.
variation_id: int, id for dialect (indexed from 0).
video_id: str, a unique video identifier.

Please be kindly advised that if you decode with different FPS, you may need to recalculate the frame_start and frame_end to get correct video segments.

Constituting subsets

As described in the paper, four subsets WLASL100, WLASL300, WLASL1000 and WLASL2000 are constructed by taking the top-K (k=100, 300, 1000 and 2000) glosses from the WLASL_vx.x.json file.

Training and Testing

I3D

cd WLASL
mkdir data

put all the videos under data/.

cp WLASL2000 -r data/

To train models, first download I3D weights pre-trained Kinetics and unzip it. You should see a folder I3D/weights/.

python train_i3d.py

To test pre-trained models, first download WLASL pre-trained weights and unzip it. You should see a folder I3D/archived/.

python test_i3d.py

By default the script tests WLASL2000. To test other subsets, please change line 264, 270 in test_i3d.py properly.

A previous release can be found here.

Pose-TGCN

Download splits file and body keypoints. Unzip them into WLASL/data. You should see WLASL/data/splits and WLASL/data/pose_per_individual_videos folders.

To train the model, modify paths in train_tgcn.py main() to point to WLASL root.

python train_tgcn.py

To test the model, first download pre-trained models and unzip to code/TGCN/archived. Then run

python test_tgcn.py

License

Licensed under the Computational Use of Data Agreement (C-UDA). Plaese refer to C-UDA-1.0.pdf for more information.

Disclaimer

All the WLASL data is intended for academic and computational use only. No commercial usage is allowed. We highly respect copyright and privacy. If you find WLASL violates your rights, please contact us.

Citation

Please cite the WLASL paper if it helps your research:

 @inproceedings{li2020word,
    title={Word-level Deep Sign Language Recognition from Video: A New Large-scale Dataset and Methods Comparison},
    author={Li, Dongxu and Rodriguez, Cristian and Yu, Xin and Li, Hongdong},
    booktitle={The IEEE Winter Conference on Applications of Computer Vision},
    pages={1459--1469},
    year={2020}
 }

Please consider citing our work on WLASL.

@inproceedings{li2020transferring,
 title={Transferring cross-domain knowledge for video sign language recognition},
 author={Li, Dongxu and Yu, Xin and Xu, Chenchen and Petersson, Lars and Li, Hongdong},
 booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
 pages={6205--6214},
 year={2020}
}

Other works you might be interested in.

@article{li2020tspnet,
  title={Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation},
  author={Li, Dongxu and Xu, Chenchen and Yu, Xin and Zhang, Kaihao and Swift, Benjamin and Suominen, Hanna and Li, Hongdong},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  pages={12034--12045},
  year={2020}
}

@inproceedings{li2022transcribing,
  title={Transcribing natural languages for the deaf via neural editing programs},
  author={Li, Dongxu and Xu, Chenchen and Liu, Liu and Zhong, Yiran and Wang, Rong and Petersson, Lars and Li, Hongdong},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={11},
  pages={11991--11999},
  year={2022}
}

jmintb/WLASL