how2sign-resnet50-mediapipe-30-pose
收藏资源简介：
How2Sign ResNet50 + MediaPipe 30-Pose Features 是一个从原始 How2Sign 语料库派生的特征数据集，专门为手语翻译和相关多模态序列建模任务设计。该数据集包含预计算的 ResNet50 视觉特征和 MediaPipe 基于姿势的特征，适用于训练和评估流程。当前上传的压缩分割档案包括训练集和验证集的前视图和侧视图特征。数据集以 .tar.zst 格式分发，旨在高效存储和传输大量预计算特征文件。该数据集适用于非商业研究用途，如手语翻译模型训练、基于特征的多模态模型训练以及避免重复特征提取。使用时应遵循原始 How2Sign 数据集的 CC BY-NC 4.0 许可条款，并正确引用原始作者的工作。
How2Sign ResNet50 + MediaPipe 30-Pose Features is a feature dataset derived from the original How2Sign corpus, specifically designed for sign language translation and related multimodal sequence modeling tasks. This dataset contains pre-computed ResNet50 visual features and MediaPipe-based pose features, which are suitable for training and evaluation workflows. The currently uploaded compressed split archive includes front-view and side-view features for both the training and validation sets. The dataset is distributed in .tar.zst format to enable efficient storage and transmission of large volumes of pre-computed feature files. This dataset is intended for non-commercial research use cases, such as sign language translation model training, feature-based multimodal model training, and avoiding redundant feature extraction. When utilizing this dataset, users must comply with the CC BY-NC 4.0 license terms of the original How2Sign dataset and properly cite the original authors' work.
How2Sign ResNet50 + MediaPipe 30-Pose Features 数据集概述
数据集简介
这是一个从原始 How2Sign 语料库构建的衍生特征数据集。它打包了针对选定How2Sign数据分割和摄像机视角预计算的ResNet50视觉特征以及MediaPipe姿态特征。
语言
- 英语
许可证
- CC BY-NC 4.0
标签
- asl
- sign-language
- how2sign
- video
- multimodal
- feature-extraction
- resnet50
- mediapipe
- pose-estimation
数据集内容
当前上传包含以下压缩分割归档文件（.tar.zst格式）：
how2sign_train_frontal_features.tar.zsthow2sign_train_side_features.tar.zsthow2sign_val_frontal_features.tar.zst
预期用途
该数据集旨在用于研究和教育用途，尤其适用于：
- 训练手语翻译模型
- 训练基于特征的多模态模型
- 避免在原始How2Sign视频上重复进行特征提取
- 使用固定的预计算输入进行可重复的实验
源数据集
原始数据来源于 How2Sign，这是一个由Duarte等人（CVPR 2021）提出的大规模多模态、多视角连续美国手语数据集。
重要说明
- 该仓库包含衍生特征，而非原始的How2Sign视频。
- 文本标注、元数据表和训练脚本不一定包含在内，除非单独上传。
- 用户应确保其使用方式符合原始How2Sign的许可证和使用条款。
引用
原始How2Sign论文
bibtex @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} }
