MagicData
收藏资源简介:
MagicData是一个大规模的轨迹控制视频数据集,由复旦大学和微软亚洲研究院创建。该数据集包含51,000个视频样本,每个样本都标注有<视频,文本,轨迹>三元组。数据集通过一个综合的数据管道进行构建,使用大型语言模型提取视频中的主要移动对象,并使用Segment Anything Model (SAM2)进行对象分割和边界框提取。该数据集旨在为轨迹控制视频生成提供高质量的数据支持,并解决视频生成中对象运动和轨迹控制精度的问题。
MagicData is a large-scale trajectory-controlled video dataset developed by Fudan University and Microsoft Research Asia. This dataset comprises 51,000 video samples, each annotated with a <video, text, trajectory> triple. It is built through a comprehensive data pipeline, in which large language models (LLMs) are utilized to extract the primary moving objects from videos, while the Segment Anything Model (SAM2) is adopted for object segmentation and bounding box extraction. The dataset is designed to provide high-quality data support for trajectory-controlled video generation, and address the issues of object motion and trajectory control accuracy in video generation.

- 1MagicMotion: Controllable Video Generation with Dense-to-Sparse Trajectory Guidance复旦大学 · 2025年



