MovieTection

该数据集包含了14,000帧与详细描述配对的基准数据，这些帧选自模型训练截止日期前后上映的电影。该数据集用于评估模型在识别版权内容方面的表现，其中包括了不同情境下的变化以及与人类参与者的表现对比。通过这一系列的实验，旨在评估视觉语言模型在检测版权内容方面的有效性。

This dataset comprises 14,000 frames of benchmark data paired with detailed descriptions, with the frames sourced from films released around the model's training cutoff date. This dataset is utilized to evaluate models' performance in copyrighted content recognition, encompassing variations across diverse scenarios and comparisons against the performance of human participants. A series of experiments conducted using this dataset aim to assess the efficacy of vision-language models in detecting copyrighted content.