sodonne6/how2sign-resnet50-mediapipe-30-pose

How2Sign ResNet50 + MediaPipe 30-Pose Features

This repository contains pre-extracted feature archives derived from the How2Sign dataset for sign language translation and related multimodal sequence modeling tasks.

The current upload includes compressed split archives in .tar.zst format:

• —how2sign_train_frontal_features.tar.zst
• —how2sign_train_side_features.tar.zst
• —how2sign_val_frontal_features.tar.zst

These archives are intended to serve as downloadable preprocessed inputs for training and evaluation pipelines, rather than as a row-based Hugging Face datasets table.

Dataset Summary

This is a derived feature dataset built from the original How2Sign corpus. It packages precomputed ResNet50 visual features together with MediaPipe pose-based features for selected How2Sign splits and camera views.

The dataset is meant for workflows such as:

• —sign language translation
• —sign language recognition
• —multimodal sequence modeling
• —feature-based training pipelines
• —experiments comparing frontal and side views

Source Dataset

The original data comes from How2Sign, a large-scale multimodal and multiview continuous American Sign Language (ASL) dataset introduced by Duarte et al. (CVPR 2021).

If you use this repository, please also cite the original How2Sign dataset paper.

What is Included

This release currently contains the following archives:

• —Train / frontal view feature archive
• —Train / side view feature archive
• —Validation / frontal view feature archive

Each archive contains preprocessed feature files extracted from the corresponding split/view of the source videos.

Format

Files are distributed as compressed tar archives:

• —format: .tar.zst

This format is intended for efficient storage and transfer of many precomputed feature files.

Intended Use

This repository is intended for research and educational use, especially for:

• —training sign language translation models
• —training feature-based multimodal models
• —avoiding repeated feature extraction on the raw How2Sign videos
• —reproducible experiments using fixed precomputed inputs

Notes

• —This repository contains derived features, not the raw How2Sign videos.
• —Text annotations, metadata tables, and training scripts are not necessarily included unless uploaded separately.
• —Users should ensure their use remains consistent with the original How2Sign license and terms of use.

License

This repository is derived from the How2Sign dataset.

The original How2Sign dataset is released under CC BY-NC 4.0 for non-commercial research use. Please consult the original dataset terms and provide proper attribution to the original authors.

Citation

Original How2Sign paper

@InProceedings{Duarte_2021_CVPR,
    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},
    title     = {How2Sign: A Large-Scale Multimodal Dataset for Continuous American Sign Language},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {2735-2744}
}