EmoCap100K
收藏资源简介:
EmoCap100K是一个大规模的、语义丰富的面部情感描述数据集,包含超过10万个样本,具有丰富和结构化的语义描述,捕捉到整体情感状态和细微的面部行为。该数据集的创建旨在解决现有面部情感识别系统在表达和适用性上的局限性,通过利用自然语言提供的灵活性和可解释性来提高系统的通用性和实用性。数据集的创建过程包括从电影中提取面部图像,并利用多模态大型语言模型(MLLMs)进行自动标注,以生成全面且结构化的情感描述。EmoCap100K数据集可用于提高面部情感理解的研究,并为学习面部情感表示提供有价值的资源。
EmoCap100K is a large-scale, semantically rich facial emotion description dataset containing over 100,000 samples, with comprehensive and structured semantic descriptions that capture both overall emotional states and subtle facial behaviors. This dataset is developed to address the limitations in expression and applicability of existing facial emotion recognition systems, and improve the generality and practicality of such systems by leveraging the flexibility and interpretability provided by natural language. The construction process of the dataset involves extracting facial images from movies, and utilizing multi-modal large language models (MLLMs) for automatic annotation to generate comprehensive and structured emotional descriptions. The EmoCap100K dataset can be used to advance research in facial emotion understanding, and serves as a valuable resource for learning facial emotion representations.
EmoCapCLIP 数据集概述
📌 基本信息
- 数据集名称: EmoCapCLIP
- 论文标题: Learning Transferable Facial Emotion Representations from Large-Scale Semantically Rich Captions
- 论文地址: https://arxiv.org/abs/2507.21015
- 作者: Licai Sun∗, Xingxun Jiang∗, Haoyu Chen, Yante Li, Zheng Lian, Biu Liu, Yuan Zong, Wenming Zheng†, Jukka M. Leppänen, Guoying Zhao†
- 机构: University of Oulu, Southeast University, University of Turku, Institute of Automation, Chinese Academy of Sciences
📝 数据集简介
- 数据集名称: EmoCap100K
- 规模: 超过100,000个样本
- 特点:
- 包含丰富且结构化的语义描述
- 捕捉全局情感状态和细粒度局部面部行为
- 提供自然语言描述作为监督信号
🎯 研究目标
- 解决当前面部情绪识别系统的局限性:
- 过度简化情绪表达为固定类别或抽象维度值
- 缺乏对丰富情绪谱系的捕捉能力
- 利用自然语言的灵活性和表达力:
- 提供更广泛、更丰富的监督来源
- 提高情绪表示的泛化能力和适用性
🏗️ 技术框架
- 方法名称: EmoCapCLIP
- 关键技术:
- 联合全局-局部对比学习框架
- 跨模态引导的正样本挖掘模块
- 全面利用多级标题信息
- 适应紧密相关表达之间的语义相似性
📊 评估结果
- 测试范围: 超过20个基准测试
- 覆盖任务: 5种不同任务
- 性能表现: 展示了优越的性能
📅 发布计划
- 代码状态: 即将发布
- 联系方式: licai.sun@oulu.fi
📚 引用信息
bibtex @article{sun2025learning, title={Learning Transferable Facial Emotion Representations from Large-Scale Semantically Rich Captions}, author={Licai Sun, Xingxun Jiang, Haoyu Chen, Yante Li, Zheng Lian, Biu Liu, Yuan Zong, Wenming Zheng, Jukka M. Leppänen, Guoying Zhao}, journal={arXiv preprint arXiv:2507.21015}, year={2025} }




