RegVID300k
收藏资源简介:
RegVID300k是由NVIDIA和延世大学联合创建的大规模区域级视频指令数据集,旨在支持视频和图像的区域级理解任务。该数据集包含98k个独特的视频,214k个区域,以及294k条区域级指令样本,数据来源于10个公开视频数据集。数据集通过自动化流水线生成,基于GPT4o的强大能力进行样本筛选和标注。RegVID300k的应用领域包括视觉常识推理、视频字幕生成和指代表达理解等任务,旨在提升多模态大语言模型在区域级理解任务中的表现。
RegVID300k is a large-scale regional video instruction dataset co-created by NVIDIA and Yonsei University, designed to support regional-level understanding tasks for videos and images. The dataset includes 98k unique videos, 214k regions, and 294k regional-level instruction samples, sourced from 10 publicly available video datasets. Generated through an automated pipeline, the dataset leverages the powerful capabilities of GPT4o for sample selection and annotation. The applications of RegVID300k encompass tasks such as visual common-sense reasoning, video captioning, and pointer expression understanding, aiming to enhance the performance of multimodal large language models in regional-level understanding tasks.




