FunQA

该数据集名为FunQA，是一个具有挑战性的视频问答（QA）数据集，旨在评估和提高基于反直觉和趣味性视频的视频推理能力。它包含三个子集：幽默QA、创意QA和魔法QA，覆盖了各种类型的令人惊讶的视频。FunQA包含的任务包括反直觉时间戳定位、详细描述以及对幽默、创造性和反直觉性的推理。该数据集规模宏大，包含4,365个视频片段和311,950组问题-答案对，其任务主要集中在视频问答领域。The dataset named FunQA is a challenging video question answering (QA) dataset designed to evaluate and enhance video reasoning capabilities based on counterintuitive and entertaining videos. It comprises three subsets: Humorous QA, Creative QA, and Magical QA, covering various types of surprising videos. The tasks included in FunQA cover counterintuitive timestamp localization, detailed description, and reasoning about humor, creativity, and counterintuitiveness. This large-scale dataset contains 4,365 video clips and 311,950 question-answer pairs, with its tasks primarily focused on the video QA domain.

搜集汇总

背景与挑战

背景概述

FunQA是一个视频问答数据集，旨在评估和增强AI模型对反直觉视频的理解能力，包括幽默、创意和魔术视频。它包含312K自由文本问答对，基于4.3K个视频片段，总时长24小时，涵盖时间戳定位、详细描述和推理等任务，并提供了扩展数据集和评估代码以支持模型开发。