VideoRef45K

该数据集包含了45,000个视频问答对，专为训练模型在基于视频的参照理解方面的能力而建立。该数据集被用于训练Artemis模型，对于评估基于视频的参照表达能力至关重要。总之，这项任务聚焦于视频基础的参照任务。This dataset contains 45,000 video question-answer pairs, which is specifically designed to train models' capabilities in video-based referential comprehension. This dataset is employed for training the Artemis model and plays a critical role in evaluating video-based referential expression capabilities. Overall, this task centers on video-grounded referential tasks.

搜集汇总

背景与挑战

背景概述

VideoRef45K是一个新建立的视频指代理解数据集，包含45,000个视频-问答对，用于训练多模态大语言模型（如Artemis）进行精细的视频指代理解，即根据视频帧中的边界框和自然语言问题描述被指代目标。