vrg-prague/OCHuman-Pose

OCHuman-Pose是原始OCHuman数据集的注释扩展，专门用于评估拥挤和严重遮挡场景下的人体姿态估计。该数据集不添加新图像，仅基于原始OCHuman图像增加和重组注释，以解决原始数据集中未标注可见人物导致假阳性问题。它提供COCO格式的边界框和关键点注释（使用标准17个关键点布局，包括鼻、眼、耳、肩、肘、腕、髋、膝、踝等），但不包含分割掩码。数据集主要用于评估，包括验证集（2,500张图像，6,546个姿态实例）和测试集（2,231张图像，5,863个姿态实例），总计4,731张图像和12,409个姿态实例，比原始OCHuman增加了超过50%的姿态实例。注释过程由专业标注员完成，采用双标注和质量检查。数据集适用于评估2D人体姿态估计、拥挤场景姿态估计、遮挡和人际交互分析，以及比较自上而下、自下而上、无检测器和迭代姿态估计方法，但不适用于训练大型模型、分割评估或掩码检测。使用前需从原始OCHuman来源下载图像，并与提供的注释文件结合。OCHuman-Pose is an annotation extension of the original OCHuman dataset, specifically designed for evaluating human pose estimation in crowded and severely occluded scenarios. It does not add new images, but only augments and reorganizes annotations based on the original OCHuman images to address the false positive issue caused by unlabeled visible individuals in the original dataset. It provides COCO-formatted bounding box and keypoint annotations using the standard 17-keypoint layout, which includes nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles, and other related body parts, but does not include segmentation masks. The dataset is mainly used for evaluation purposes, including a validation set (2,500 images, 6,546 pose instances) and a test set (2,231 images, 5,863 pose instances), with a total of 4,731 images and 12,409 pose instances, representing an increase of over 50% in pose instances compared to the original OCHuman dataset. The annotation process was completed by professional annotators via double annotation and quality inspection procedures. This dataset is suitable for evaluating 2D human pose estimation, pose estimation in crowded scenes, occlusion and interpersonal interaction analysis, as well as comparing top-down, bottom-up, detector-free, and iterative pose estimation methods. However, it is not applicable for training large-scale models, segmentation evaluation, or mask detection. Prior to use, images must be downloaded from the original OCHuman dataset source and combined with the provided annotation files.