VideoCoT, TopicQA, TopicCoT
收藏资源简介:
VideoCoT、TopicQA和TopicCoT是由华南理工大学、字节跳动和琶洲实验室联合创建的视频链式思考(CoT)数据集。这些数据集包含11,000条视频,每条视频对应2个问题,总计22,000个问题,数据来源于Kinetics-700。数据集的创建过程结合了自动标注和人工校验,旨在通过主动学习范式提高标注效率和数据质量。这些数据集主要应用于视频开放式问答和提升多模态大语言模型的推理能力,特别是在视频理解和复杂语义捕捉方面。
VideoCoT, TopicQA and TopicCoT are video chain-of-thought (CoT) datasets jointly created by South China University of Technology, ByteDance and Pazhou Laboratory. These datasets include 11,000 videos, each corresponding to two questions, totaling 22,000 questions, with all data sourced from Kinetics-700. The construction of these datasets combines automatic annotation and manual verification, aiming to improve annotation efficiency and data quality through the active learning paradigm. These datasets are mainly applied to video open-ended question answering and enhancing the reasoning capabilities of multimodal large language models, particularly in video understanding and complex semantic capture.

- 1VideoCoT: A Video Chain-of-Thought Dataset with Active Annotation Tool华南理工大学,广州,中国;字节跳动,北京,中国;琶洲实验室,广州,中国 · 2024年



