Fake-Motion Dataset

Fake-Motion Dataset是由莫斯科国立大学的研究团队创建的一个用于视频人物抠图训练的数据集。该数据集包含40,074张带有真实alpha matte的前景人物图像和27个背景视频，用于模拟真实视频中的运动。创建过程中，研究团队采用了一种新颖的假运动算法，通过扭曲前景图像来生成训练视频片段，从而在没有大型标注视频数据集的情况下训练深度神经网络。该数据集主要应用于视频处理领域，旨在解决视频中人物抠图的透明度映射问题，提高视频抠图的时序稳定性。

The Fake-Motion Dataset is a dataset for video human matting training, developed by a research team from Lomonosov Moscow State University. This dataset contains 40,074 foreground human images with ground-truth alpha mattes and 27 background videos, which are utilized to simulate motions in real-world videos. During the creation process, the research team adopted a novel fake motion algorithm to generate training video clips by warping the foreground images, thereby enabling the training of deep neural networks without large-scale annotated video datasets. This dataset is primarily applied in the field of video processing, with the goal of solving the transparency mapping issue of human matting in videos and enhancing the temporal stability of video matting.