M-VAD Names

M-VAD Names数据集是由恩佐·法拉利工程学院，摩德纳与雷焦艾米利亚大学创建，旨在支持视频字幕架构的命名能力开发。该数据集包含63,442个视觉轨迹和34,388个文本提及，均与角色身份关联。通过半自动标注过程，数据集精确地标注了角色的视觉外观，并手动纠正了原始M-VAD标注中的错误。此数据集不仅用于视频字幕，还可用于动作识别和视频的视觉-语义空间训练，旨在解决现有视频字幕模型无法正确提及角色名称的问题。

The M-VAD Names dataset was developed by the Enzo Ferrari School of Engineering, University of Modena and Reggio Emilia, with the goal of supporting the development of naming capabilities for video captioning architectures. This dataset contains 63,442 visual trajectories and 34,388 textual mentions, all associated with character identities. Through a semi-automatic annotation pipeline, the dataset accurately annotates the visual appearance of characters and manually corrects errors in the original M-VAD annotations. This dataset can be applied not only to video captioning, but also to action recognition and visual-semantic space training for videos, and is designed to address the issue that existing video captioning models fail to correctly mention character names.