人类姿势

MPII 人体姿势描述数据集是对广泛使用的 MPII 人体姿势数据集的扩展，添加了丰富的文本注释，这些注释由多种先进语言模型（LLM）生成，包含对正在执行的活动、在场人数及特定姿势的详细描述；数据集与 MMPose 中提供的图像分割一致，包含 14644 个训练样本和 2723 个验证样本，每张图像附有不同 LLM 生成的一个或多个姿势描述，还包含源自原始数据集注释的活动类型、人数、姿势关键点等额外信息；其基于 BSD-2-Clause 许可证，支持免费使用、修改和分发，需适当归属The MPII Human Pose Description Dataset is an extension of the widely adopted MPII Human Pose Dataset, which adds rich textual annotations generated by multiple state-of-the-art Large Language Models (LLMs). These annotations include detailed descriptions of ongoing activities, the number of people present, and specific poses. The dataset shares the same image splits as those provided in MMPose, containing 14,644 training samples and 2,723 validation samples. Each image is attached with one or more pose descriptions generated by different LLMs, and also incorporates additional information such as activity types, number of people, and pose keypoints derived from the annotations of the original dataset. It is licensed under the BSD-2-Clause license, permitting free use, modification and distribution with proper attribution.

搜集汇总

背景与挑战

背景概述

该数据集是MPII人体姿势数据集的扩展版本，添加了由GPT-4、GPT-3.5-turbo等先进语言模型生成的丰富文本注释，包含14644个训练样本和2723个验证样本，以JSON格式组织。它适用于多模态学习、细粒度活动识别和视觉-文本关联研究，基于BSD-2-Clause许可证，支持免费使用和修改。