Javelin_Throw_Dataset
收藏资源简介:
本研究提出了一种基于生物力学相位的运动视频时间分割新框架,并通过结构化最优传输概念增强知名的关注点时空图卷积网络。通过该方法,无需昂贵的标签,即可识别运动相位转换。实验结果表明,该方法在测试数据上取得了71.02%的平均平均精度(mAP)和74.61%的F1分数,显著高于竞争基线。此外,还发布了一个包含211个手动标注的专业标枪投掷视频数据集,这些视频来自真实世界的比赛,涵盖了关键生物力学相位:步进、驱动、投掷和恢复。该数据集可用于训练和评估无监督分割模型,以进行细粒度的运动分析。
This study proposes a novel framework for temporal segmentation of motion videos based on biomechanical phases, which enhances the well-known attention-based spatial-temporal graph convolutional network via the concept of structured optimal transport. This method enables identification of motion phase transitions without costly labeled data. Experimental results demonstrate that the proposed method achieves a mean Average Precision (mAP) of 71.02% and an F1-score of 74.61% on the test dataset, significantly outperforming competing baseline models. In addition, a manually annotated dataset consisting of 211 professional javelin throw videos sourced from real-world competitions is publicly released, covering the key biomechanical phases: stepping, driving, throwing, and recovery. This dataset can be utilized to train and evaluate unsupervised segmentation models for fine-grained motion analysis.
Javelin Throw Dataset 概述
数据集简介
Javelin Throw Action Dataset 包含同步的视频文件、骨骼姿态数据和标注动作阶段数据,用于标枪投掷分析。该数据集旨在支持体育动作分割、姿态估计和阶段检测的研究。
数据集组成
视频文件
- 格式:MP4
- 命名规则:vid_<index>.mp4(例如 vid_0.mp4, vid_1.mp4)
- 内容:原始标枪投掷视频
- 下载地址:https://drive.google.com/file/d/1fKfZNCHdr3NK7Ppnfcb3oV-57pTRH9ys/view?usp=sharing
骨骼数据
- 格式:JSON
- 命名规则:video_<index>.json(与视频文件对应)
- 内容:使用 MMPose 从视频中提取的 2D 姿态骨骼数据
- 下载地址:https://drive.google.com/file/d/1DM9kKhr7CwC61Df60mzRVP6s_bcI8XP8/view?usp=sharing
标注数据
- 格式:JSON
- 命名规则:vid_<index>.json
- 内容:标枪投掷阶段的帧级标注,包括 Steps、Drive、Throw、Recovery 四个阶段
- 下载地址:https://drive.google.com/file/d/1aaG10_POwS2g9pUnqjgpQrB1Kp9Htr9L/view?usp=sharing

- 1Biomechanical-phase based Temporal Segmentation in Sports Videos: a Demonstration on Javelin-Throw印度理工学院甘地讷格尔 · 2025年



