3D Shot Posture (3DSP) dataset

该数据集名为3DSP，对于开发AutoSoccerPose技术至关重要，它作为在专业足球比赛直播视频中2D姿态估计的基准。该数据集包含了一系列捕捉足球动作的图像，重点关注射门等不同的动作。此外，数据集不仅包含了2D姿态标注，还包含了估计的3D姿态，训练集由每项射门动作的20张裁剪图像组成，而测试集则仅包含轨迹信息。该数据集的规模涵盖了4000张标注图像，这些图像基于200个射门动作，来自22场独特的比赛。其任务是对足球动作进行2D和3D姿态估计以及姿势分析。This dataset, named 3DSP, is critical for the development of AutoSoccerPose technology and serves as a benchmark for 2D pose estimation in live broadcasts of professional soccer matches. It contains a series of images capturing soccer actions, with a focus on diverse movements such as shots. Additionally, the dataset includes both 2D pose annotations and estimated 3D poses. The training set consists of 20 cropped images for each shot, while the test set only contains trajectory information. The dataset spans 4,000 annotated images derived from 200 shooting actions across 22 unique matches. The task focuses on 2D and 3D pose estimation and pose analysis for soccer actions.

搜集汇总

背景与挑战

背景概述

3D Shot Posture (3DSP) dataset是一个专注于足球射门动作分析的3D姿态数据集，源自CVPR 2024论文“AutoSoccerPose”。该数据集包含训练集（20张裁剪图像×200个射门）和测试集（20张裁剪图像×10个射门），提供轨迹、2D和3D关键点数据，用于自动化姿态提取和聚类分析。数据来源于SoccerNet广播视频，通过YOLO v8和BoT-Sort处理，并结合手动标注与MotionAGFormer模型生成3D姿态，适用于计算机视觉和体育分析研究。