RealFlow

RealFlow是一个基于期望最大化（EM）框架的大型光学流数据集，由清华大学软件学院和Megvii Technology共同创建。该数据集直接从真实视频中生成，包含161709条数据，旨在解决传统光学流数据集生成中的域差异问题。RealFlow通过迭代学习过程，逐步提高光学流网络的性能和数据集的质量。数据集的应用领域包括提高光学流估计的准确性和泛化能力，特别是在真实世界场景中的应用。

RealFlow is a large-scale optical flow dataset based on the Expectation-Maximization (EM) framework, jointly created by the School of Software, Tsinghua University and Megvii Technology. This dataset is directly generated from real-world videos, containing 161,709 data samples. It aims to solve the domain shift problem encountered in the generation of traditional optical flow datasets. RealFlow improves the performance of optical flow networks and the quality of the dataset itself through an iterative learning process. Its application fields include enhancing the accuracy and generalization ability of optical flow estimation, especially for real-world scene applications.