FS-Jump3D Dataset
收藏资源简介：
名古屋大学的研究团队精心打造了FS-Jump3D数据集，旨在通过光学无标记运动捕捉技术，细致记录花样滑冰中复杂而动态的跳跃动作。该数据集不仅填补了研究领域中针对花样滑冰三维姿态数据的空白，还提出了一种新的细粒度标注方法，使动作分割模型能够学习跳跃过程。FS-Jump3D数据集包含12个视角的视频和86个关键点的三维跳跃姿态，为理解花样滑冰跳跃提供了宝贵的资源。通过与现有的三维姿态数据集相结合，FS-Jump3D进一步推动了体育领域三维姿态估计技术的发展。数据集创建过程中，研究者在冰面上布置了12台高速摄像机，以60fps的帧率捕捉了包括三周跳在内的高难度跳跃动作。数据集的标注工作细致入微，将跳跃动作分为“起跳”、“跳跃中”和“着陆”三个阶段，为模型提供了精确的时空语义信息。此外，研究还验证了3D姿态特征作为输入数据在花样滑冰动作分割任务中的有效性。
The research team at Nagoya University has meticulously developed the FS-Jump3D dataset, aiming to record the complex and dynamic jump movements in figure skating via optical markerless motion capture technology. This dataset not only fills the gap in 3D pose data for figure skating in the research community, but also proposes a novel fine-grained annotation method that enables action segmentation models to learn the entire jump process. The FS-Jump3D dataset includes 12-view videos and 3D jump poses of 86 key points, providing a valuable resource for the study of figure skating jumps. Combined with existing 3D pose datasets, FS-Jump3D further advances the development of 3D pose estimation technologies in the sports domain. During the dataset creation phase, the researchers deployed 12 high-speed cameras on the ice rink to capture high-difficulty jump movements including triple jumps at a frame rate of 60fps. The annotation work for the dataset is highly meticulous, dividing jump movements into three stages: takeoff, mid-air, and landing, providing models with precise spatiotemporal semantic information. Additionally, the study verified the effectiveness of 3D pose features as input data for figure skating action segmentation tasks.
FS-Jump3D 数据集
概述
FS-Jump3D 数据集是首个包含3D姿态数据和12个视角视频数据的花样滑冰跳跃数据集。该数据集通过无标记运动捕捉系统（Theia3D, Theia）和12台高速摄像机（Miqus Video, Qualisys）在冰上滑冰场中捕捉跳跃数据。
数据内容
- c3d: 302.6 MB
- json: 505.2 MB
- videos: 8.84 GB
下载
数据集可通过以下链接下载：
- c3d: GoogleDrive
- json: GoogleDrive
- videos: GoogleDrive
使用方法
使用C3D文件
可通过QTM（Qualisys Track Manager）软件打开和可视化原始C3D文件，并导出为多种格式。
使用JSON文件
JSON文件包含与原始C3D文件相同的数据。可通过以下命令将JSON文件转换为常见的3D姿态格式： zsh python utils/format.py
许可证
FS-Jump3D 数据集采用 Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) 许可证。
引用
@inproceedings{MMSports2024_tanaka, author = {Tanaka, Ryota and Suzuki, Tomohiro and Fujii, Keisuke}, title = {3D Pose-Based Temporal Action Segmentation for Figure Skating: A Fine-Grained and Jump Procedure-Aware Annotation Approach}, booktitle = {Proceedings of the 7th ACM International Workshop on Multimedia Content Analysis in Sports}, series = {MMSports 24}, year = {2024}, isbn = {979-8-4007-1198-5/24/10}, location = {Melbourne, VIC, Australia}, pages = {1--10}, numpages = {10}, doi = {10.1145/3689061.3689077}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, keywords = {Temporal action segmentation, Human pose estimation, Sports, Datasets, Annotation, Computer vision}, }
