CCTV Human Pose Estimation Dataset

This dataset contains 700 images from independent CCTV view points, with the intended use being to test the performance of human pose estimation systems on CCTV data. Generally, such systems are trained on high quality, non CCTV data. This dataset enabled the evaluation of these systems as part of my PhD Project, of which a noticeable drop in performance was found compared to the performance found on other pose estimation benchmarks. Naturally, the dataset is quite small and so should probably be expanded upon in the future, possibly to enable training in addition to evaluation.

Annotations were manually done and validated by comparing a small set of annotations to two other annotators.
These annotations are set up according to the MSCOCO format and its evaluation toolkit.

本数据集汇聚了700张独立监控摄像头视角的图像，旨在检验人类姿态估计系统在监控数据上的性能。通常，此类系统系基于高质量的非监控数据进行训练。本数据集的构建旨在作为本人博士研究项目的一部分，用以评估相关系统性能，研究发现其性能相较于其他姿态估计基准测试存在显著下降。显然，该数据集规模相对较小，未来可能需要进一步扩充，以便实现除评估之外的训练目的。
标注工作系由人工完成，并通过将少量标注与两位其他标注者进行比较进行验证。这些标注遵循了MSCOCO格式及其评估工具包的设置。