EHPAUB
收藏资源简介:
EHPAUB是一个包含200,328个样本的数据集,旨在为以人为中心的任务微调模型,重点关注对话、详细描述和复杂推理三个方面。数据集由东京电机大学信息学部和TikTok Inc合作创建,用于提升多模态模型在理解人类姿态和动作方面的能力。数据集的创建过程包括将人类关键点与传统的视觉特征(如标题和边界框)相结合,以生成更具精确性的数据。该数据集应用于评估模型在理解人类姿态和动作方面的性能,并在辅助机器人、医疗保健和人机交互等领域具有潜在的应用价值。
EHPAUB is a dataset consisting of 200,328 samples, designed to fine-tune models for human-centric tasks, with three core focuses: dialogue, detailed description, and complex reasoning. It was co-created by the Faculty of Informatics, Tokyo Denki University and TikTok Inc., with the objective of enhancing multimodal models' capability to understand human postures and movements. The dataset construction process integrates human keypoints with conventional visual features including captions and bounding boxes to generate more precise data. This dataset is employed to evaluate models' performance in understanding human postures and movements, and holds potential application value across domains such as assistive robotics, healthcare, and human-computer interaction.
LLaVA-Pose数据集概述
数据集简介
- 提供生成关键点集成指令跟随数据的方法,用于增强多模态模型对人体姿态和动作的理解
- 基于LLaVA框架构建
- 相关研究论文:Keypoint-Integrated Instruction-Following Data Generation for Enhanced Human Pose and Action Understanding in Multimodal Models
数据结构
plaintext Keypoint-Instruction-Tuning/ ├── data_generation/ │ ├── conversation_gen.py │ ├── detailed_description_gen.py │ └── complex_reasoning_gen.py ├── datasets/ │ ├── generated_data_conversation.json │ ├── generated_data_detailed.json │ └── generated_data_reasoning.json ├── LLaVA/ │ └── [LLaVA original files here]
数据生成方法
-
对话数据生成 bash python data_generation/conversation_gen.py
-
详细描述数据生成 bash python data_generation/detailed_description_gen.py
-
复杂推理数据生成 bash python data_generation/complex_reasoning_gen.py
安装要求
bash pip install -r requirements.txt
数据存储
- 所有生成数据保存在
datasets/目录下 - 包含三种JSON格式数据文件:
- generated_data_conversation.json
- generated_data_detailed.json
- generated_data_reasoning.json
相关资源
- 原始LLaVA框架:LLaVA repository

- 1LLaVA-Pose: Enhancing Human Pose and Action Understanding via Keypoint-Integrated Instruction Tuning东京电机大学信息学部 · 2025年



