PFED5+
收藏资源简介:
PFED5+是一个专门用于面部表情质量评估(FEQA)的扩展数据集,由布里斯托大学和兰卡斯特大学的研究团队联合构建。该数据集在原始PFED5的基础上,新增了专家指导的、临床审查过的视频片段级运动描述文本注释,共计包含2,811个视频片段,数据来源于帕金森病患者的特定面部动作任务。其创建过程采用了系统化的标注框架,将每个视频按时间结构划分为动作前、中、后三个阶段,并由专家定义临床相关面部区域进行精细观察和结构化描述,再经GPT-4o进行文本流畅性优化并最终由临床专家双重验证。该数据集主要应用于医疗健康领域,旨在支持生成式可解释性研究,通过联合预测临床严重程度评分并生成结构化证据报告,以解决传统FEQA方法仅输出分数而缺乏可审计运动证据的问题,从而增强神经运动障碍评估的透明度和临床实用性。
PFED5+ is an extended dataset specifically designed for Facial Expression Quality Assessment (FEQA), jointly constructed by research teams from the University of Bristol and Lancaster University. Building upon the original PFED5 dataset, it adds expert-guided, clinically reviewed video clip-level textual annotations of motion descriptions, totaling 2,811 video clips sourced from targeted facial movement tasks performed by patients with Parkinson’s disease. Its development adopted a systematic annotation framework: each video is divided into three temporal stages—pre-action, action, and post-action—with experts defining clinically relevant facial regions for fine-grained observation and structured description. The annotations then underwent text fluency optimization via GPT-4o, followed by final double validation by clinical experts. This dataset is primarily applied in the healthcare domain, aiming to support generative explainability research. It addresses the key limitation of traditional FEQA methods, which only output scores without providing auditable motor evidence, by jointly predicting clinical severity scores and generating structured evidence reports, thereby enhancing the transparency and clinical utility of neuromotor disorder assessment.
TraMP-LLaMA 数据集概述
基本信息
- 项目名称:TraMP-LLaMA: Generative Interpretability with Decoupled Instruction Tuning for Facial Expression Quality Assessment
- 论文地址:arXiv版本
- 核心任务:提出统一的多模态框架TraMP-LLaMA,能够从面部运动线索中联合预测严重程度评分并生成结构化的文本报告。
数据集:PFED5+
- 来源扩展:在PFED5数据集基础上,添加了专家指导的文本运动描述,构建了增强版PFED5+数据集。
- 视频帧和MDS-UPDRS标签获取:需通过QAFE-Net仓库申请访问权限。
- 运动描述标签:提供链接(当前为空)。
引用信息
- 若在研究中使用了本工作,建议引用相关论文,包括TraMP-LLaMA论文、Trajectory-guided Motion Perception论文以及QAFE-Net论文。
致谢
- 本代码库基于VideoLLaMA3构建。

- 1TraMP-LLaMA: Generative Interpretability with Decoupled Instruction Tuning for Facial Expression Quality Assessment布里斯托大学·计算机科学学院; 布里斯托大学·转化健康科学学院; 兰卡斯特大学·计算与通信学院 · 2026年




