HealthVidQA-CRF, HealthVidQA-Prompt

本研究创建了两个大规模的医疗视频问答数据集：HealthVidQA-CRF和HealthVidQA-Prompt，旨在通过医疗视频提供健康相关问题的视觉答案。HealthVidQA-CRF包含23,434个视频-问题-答案三元组，来自11,708个医疗视频，每个视频平均包含2个视觉答案段，时长约73.33秒。HealthVidQA-Prompt则包含52,771个三元组，来自13,990个视频，每个视频平均包含3.77个视觉答案段，时长约33.88秒。这些数据集通过自动化的管道方法生成，旨在提高医疗视觉答案定位任务的性能，并可用于预训练语言-视觉模型，以进一步增强性能。This study constructs two large-scale medical video question answering datasets: HealthVidQA-CRF and HealthVidQA-Prompt, which are designed to provide visual answers to health-related questions using medical videos. HealthVidQA-CRF consists of 23,434 video-question-answer triples sourced from 11,708 medical videos, where each video contains an average of 2 visual answer segments and has a duration of approximately 73.33 seconds. HealthVidQA-Prompt, by contrast, includes 52,771 video-question-answer triples from 13,990 videos, with each video averaging 3.77 visual answer segments and a duration of around 33.88 seconds. These datasets were generated via an automated pipeline approach, aiming to improve the performance of medical visual answer localization tasks, and can be utilized for pre-training language-vision models to further enhance their performance.

搜集汇总

背景与挑战

背景概述

HealthVidQA-CRF和HealthVidQA-Prompt是两个大规模医疗视频问答数据集，旨在通过医疗视频提供健康相关问题的视觉答案。HealthVidQA-CRF包含23,434个三元组，平均每个视频有2个视觉答案段，时长较长；HealthVidQA-Prompt包含52,771个三元组，平均每个视频有3.77个视觉答案段，时长较短。这些数据集通过自动化方法生成，用于提升医疗视觉答案定位任务性能，并支持语言-视觉模型预训练。