MADS 人体动作数据集 包含武术，舞蹈和体育数据集

Human pose estimation is one of the most popular research topics in the past two decades, especially with the introduction of human pose datasets for benchmark evaluation. These datasets usually capture simple daily life actions. Here, we introduce a new dataset, the Martial Arts, Dancing and Sports (MADS), which consists of challenging martial arts actions (Tai-chi and Karate), dancing actions (hip-hop and jazz), and sports actions (basketball, volleyball, football, rugby, tennis and badminton). Two martial art masters, two dancers and an athlete performed these actions while being recorded with either multiple cameras or a stereo depth camera. In the multi-view or single-view setting, we provide three color views for 2D image-based human pose estimation algorithms. For depth-based human pose estimation, we provide stereo-based depth images from a single view. All videos have corresponding synchronized and calibrated ground-truth poses, which were captured using a Motion Capture system. We provide initial baseline results on our dataset using a variety of tracking frameworks, including a generative tracker based on the annealing particle filter and robust likelihood function, a discriminative tracker using twin Gaussian processes, and hybrid trackers, such as Personalized Depth Tracker. The results of our evaluation suggest that discriminative approaches perform better than generative approaches when there are enough representative training samples, and that the generative methods are more robust to diversity of poses, but can fail to track when the motion is too quick for the effective search range of the particle filter. The data was recorded in a studio environment with some background clutter. The video data was recorded with Point Grey Bumblebee-II cameras. The multi-view data was collected with 3 cameras placed around the capture space, while the stereo images were collected from one viewpoint. The multi-view data was captured at 15 fps, and the cameras were synchronized automatically when connected to the same hub. The depth data (stereo image) was captured at 10 fps or 20 fps. The resolution of the images are 1024 × 768. The ground-truth pose data was captured using a MOCAP system which works at 60 fps. All videos and motion capture data are calibrated to the same coordinate and synchronized. The MADS dataset contains 5 action categories (Tai-chi, Karate, Jazz dance, Hip-hop dance, and Sports), totalling about 53,000 frames. Each action category consists of 6 sequences. Example poses are show below: We test several state-of-the-art methods on our MADS dataset, including both generative trackers and discriminative trackers, the result demos can be found on the demo link.

人体姿态估计（human pose estimation）是近二十年来最受关注的研究课题之一，尤其是随着用于基准评测的人体姿态数据集的问世，该领域的研究热度进一步提升。此类数据集通常以日常简单动作为采集目标。

本文介绍一款全新数据集——武术、舞蹈与运动数据集（Martial Arts, Dancing and Sports, 简称MADS），该数据集涵盖极具挑战性的武术动作（太极与空手道）、舞蹈动作（嘻哈与爵士舞）以及体育运动动作（篮球、排球、足球、橄榄球、网球与羽毛球）。两位武术大师、两位舞者与一名运动员完成全部动作，采集设备涵盖多相机阵列与立体深度相机两种方案。

针对多视角与单视角两种设置，我们为基于二维图像的人体姿态估计算法提供三组彩色视图数据；针对基于深度信息的人体姿态估计算法，则提供单视角下的立体深度图像。所有视频均配有经同步与标定的真实姿态标注（ground-truth poses），此类标注通过动作捕捉（Motion Capture）系统采集获取。

我们基于该数据集，结合多种跟踪框架开展了基准性能测试，其中包括基于退火粒子滤波器与鲁棒似然函数的生成式跟踪器、采用孪生高斯过程的判别式跟踪器，以及诸如个性化深度跟踪器（Personalized Depth Tracker）这类混合跟踪器。实验评估结果显示：当拥有足够多具有代表性的训练样本时，判别式方法的性能优于生成式方法；生成式方法对姿态多样性具备更强鲁棒性，但当运动速度过快超出粒子滤波器的有效搜索范围时，此类方法可能无法完成跟踪任务。

数据集采集于存在一定背景杂波的演播室环境中。视频数据采用Point Grey Bumblebee-II相机采集。多视角数据由环绕采集空间布置的三台相机采集，而立体图像仅从单一场景视点采集。多视角数据采集帧率为15 fps，当相机连接至同一集线器时可实现自动同步。深度数据（立体图像）的采集帧率为10 fps或20 fps，图像分辨率均为1024 × 768。

真实姿态标注数据通过帧率为60 fps的动作捕捉（MOCAP）系统采集。所有视频与动作捕捉数据均完成统一坐标系标定与同步处理。

MADS数据集共包含5类动作：太极、空手道、爵士舞、嘻哈舞以及体育运动，总帧数约为53000帧，每类动作均包含6个序列。示例姿态如下所示：

我们在MADS数据集上测试了多款当前主流方法，涵盖生成式与判别式跟踪器两类，相关结果演示可通过演示链接获取。

搜集汇总

背景与挑战

背景概述

MADS人体动作数据集是一个专注于武术、舞蹈和体育动作的多视角和深度图像数据集，用于3D人体姿态估计研究。数据集包含约53,000帧的同步彩色和深度图像，以及由运动捕捉系统提供的地面真实姿态数据，总大小为23.45GB。