Multi-VSL-front-skeleton
收藏资源简介:
Multi-VSL(前视图)DWPose 骨架数据集是基于 Multi-VSL 越南手语语料库前摄像头视频片段,使用 DWPose 算法提取的全身二维姿态关键点数据集。该数据集包含 28,406 个剪辑片段,每个片段存储为一个独立的 .npz 文件,总数据量约 4.6 GB。文件按照 signer 目录组织,共有 30 个 signer 目录,每个目录包含 628 至 1,167 个 .npz 文件。每个 .npz 文件包含以下字段:all_xy(形状为 (T, 128, 2) 的 float16 数组,表示每帧关键点的像素坐标)、all_score(形状为 (T, 128) 的 float16 数组,表示每个关键点的置信度)、detected(形状为 (T,) 的 int8 数组,表示该帧是否检测到人物)、frame_size(形状为 (2,) 的 int32 数组,表示源视频帧尺寸)、fps(标量 float32,表示源视频帧率)。其中 T 为剪辑的帧数,128 个关键点遵循 DWPose/COCO-WholeBody 布局,包括 17 个身体关键点、6 个脚部关键点、68 个面部关键点和 42 个手部关键点(每只手 21 个)。文件名格式为 <session>___<view>_<device>_<signer>_<view>_<order>_<clip_index>.npz,例如 01_Co-Hien_100-200_1-2-3_0118___center_device10_signer01_center_ord1_100.npz。该数据集适用于手语识别、姿态估计、人体动作分析等任务,尤其适合越南手语相关研究。
The Multi-VSL (Front View) DWPose Skeleton Dataset is a whole-body 2D pose keypoint dataset extracted from the front camera video clips of the Multi-VSL Vietnamese Sign Language Corpus using the DWPose algorithm. It contains 28,406 clips, each stored as an independent .npz file, with a total data volume of approximately 4.6 GB. The files are organized by signer directories, with 30 signer directories, each containing 628 to 1,167 .npz files. Each .npz file includes the following fields: all_xy (a float16 array of shape (T, 128, 2) representing pixel coordinates of keypoints per frame), all_score (a float16 array of shape (T, 128) representing confidence scores for each keypoint), detected (an int8 array of shape (T,) indicating whether a person is detected in the frame), frame_size (an int32 array of shape (2,) representing the source video frame dimensions), and fps (a scalar float32 representing the source video frame rate). T is the number of frames in the clip. The 128 keypoints follow the DWPose/COCO-WholeBody layout, including 17 body keypoints, 6 foot keypoints, 68 facial keypoints, and 42 hand keypoints (21 per hand). The filename format is <session>___<view>_<device>_<signer>_<view>_<order>_<clip_index>.npz, e.g., 01_Co-Hien_100-200_1-2-3_0118___center_device10_signer01_center_ord1_100.npz. This dataset is suitable for tasks such as sign language recognition, pose estimation, human action analysis, and especially for Vietnamese Sign Language research.
数据集概述
Multi-VSL (front view) — DWPose skeletons 是一个用于越南手语识别的全身2D姿态关键点数据集,从 Multi-VSL 越南手语语料库的前置摄像头视频片段中提取。
基本信息
- 许可证:unknown(未知)
- 任务类别:other(其他)
- 语言:越南语(vi)
- 标签:手语、越南手语、姿态估计、DWPose、骨架
- 数据规模:10K < n < 100K
数据内容
- 片段数量:28,406 个视频片段,每个片段对应一个
.npz文件 - 总大小:约 4.6 GB
- 目录结构:
data/<signer>/<clip>.npz,共 30 个手语者目录,每个目录包含 628 至 1,167 个文件(由于 Hugging Face 限制目录内文件数不能超过 10,000,故采用非扁平目录结构)
每个 .npz 文件包含的字段
| 字段名 | 形状 | 数据类型 | 描述 |
|---|---|---|---|
all_xy |
(T, 128, 2) |
float16 |
每帧关键点像素坐标 |
all_score |
(T, 128) |
float16 |
每个关键点的置信度 |
detected |
(T,) |
int8 |
该帧是否检测到人物(1 为是,0 为否) |
frame_size |
(2,) |
int32 |
源视频帧尺寸 |
fps |
标量 | float32 |
源视频帧率 |
其中,T 为片段的帧数。128 个关键点遵循 DWPose / COCO-WholeBody 布局,包含:17 个身体关键点 + 6 个脚部关键点 + 68 个面部关键点 + 42 个手部关键点(每只手 21 个)。
文件命名格式
文件名格式为:
<session><view><device><signer><view><order><clip_index>.npz
示例:01_Co-Hien_100-200_1-2-3_0118___center_device10_signer01_center_ord1_100.npz
使用方式
可通过 Hugging Face Hub 下载单个文件,例如:
python import numpy as np from huggingface_hub import hf_hub_download
path = hf_hub_download( "Tri1/Multi-VSL-front-skeleton", "data/signer01/01_Co-Hien_100-200_1-2-3_0118___center_device10_signer01_center_ord1_100.npz", repo_type="dataset", ) d = np.load(path) xy, score = d["all_xy"], d["all_score"] # (T, 128, 2), (T, 128)
也可下载整个数据集或仅下载某个手语者的数据:
python from huggingface_hub import snapshot_download
下载全部数据
snapshot_download("Tri1/Multi-VSL-front-skeleton", repo_type="dataset", local_dir="skeleton")
仅下载 signer01 的数据
snapshot_download("Tri1/Multi-VSL-front-skeleton", repo_type="dataset", allow_patterns="data/signer01/*", local_dir="skeleton")




