SUPERSEDED - WeightGait Dataset

## This item has been replaced by the one which can be found at https://doi.org/10.7488/ds/7897 ## Introduction: Here we introduce the WeightGait dataset: a dataset developed for the purposes of facilitating vision-based gait assessment methodologies with more realistic conditions comparable to real world use. The motivation for this dataset is to create a testing environment for gait assessment algorithms that is closer to the realities of application. To accomplish this, unlike other similar datasets, we do two main things uniquely: We simulate overlapping abnormalities, for a total of 9 different combinations of abnormality detailed below. The background and equipment used are imperfect and noisy to simulate the similar hardship experienced when trying to install a gait monitor into someone's home. This means cheap recording equipment for scalability resulting in relatively low-frames per recording. It also means slight feet/head clipping at times, only a single camera view to detect depth and no curation to the background or the clothing/walking speed of the participants. In order to preserve privacy, all original faces in the videos have been replaced by a deep-fake variation of the original created by the algorithm given in the paper 'DeepPrivacy'. Each frame has an independent new face and as a consequence, there is some flickering on the faces in the videos. The original 2D joint positions are estimated on the original videos using a lightweight implementation of the algorithm given in the paper 'HigherHRNet'.

## 本项已被可通过链接https://doi.org/10.7488/ds/7897获取的版本替换 ##

引言：本文介绍WeightGait数据集，该数据集旨在推动基于视觉的（vision-based）步态评估方法的发展，使其条件与现实世界的使用场景更为贴近。构建该数据集的核心动机是为步态评估算法搭建一个更贴近实际应用场景的测试环境。为此，与其他同类数据集相比，我们采取了两项独特设计：其一，模拟重叠异常（overlapping abnormalities），共包含下文详述的9种不同异常组合；其二，采用非理想且带噪声的背景与设备设置，以模拟家庭环境中安装步态监测设备时面临的类似挑战——例如，使用低成本录制设备导致每段记录的帧率相对较低，偶尔会出现脚部或头部被截断的情况，仅配备单摄像头（无法实现深度检测），且未对背景或参与者的服装、行走速度进行统一整理。为保护隐私，视频中所有原始人脸均已被《DeepPrivacy》论文所述算法生成的深度伪造（deep-fake）变体替换；由于每帧均采用独立生成的新面孔，视频中的人脸会出现轻微闪烁现象。原始视频中的2D关节位置（2D joint positions）是通过《HigherHRNet》论文所述算法的轻量级实现来估计的。