光学流数据集合成

本研究介绍了一种新的光学流数据集合成方法，该数据集由伯尔尼大学创建，旨在通过无监督方式生成大量具有精确光学流信息的真实图像对。数据集利用未配对的真实图像，通过模拟随机扭曲、超像素遮挡、阴影和光照变化等技术，生成具有真实纹理和合成运动的数据。该数据集的创建过程无需手动干预，适用于训练光学流估计模型，特别是在解决因遮挡、反射、纹理缺失和光照变化引起的模糊性问题。此外，数据集的应用领域包括提高计算机视觉中光学流估计的准确性和效率，尤其是在处理复杂场景和动态变化环境中的光学流问题。

This study introduces a novel optical flow dataset synthesis method. Developed by the University of Bern, this dataset aims to generate a large number of realistic image pairs with accurate optical flow information via an unsupervised approach. The dataset leverages unpaired real images, and generates data with realistic textures and synthesized motions by simulating techniques such as random warping, superpixel-based occlusion, shading and illumination variations. The creation process of this dataset requires no manual intervention, making it suitable for training optical flow estimation models, especially for resolving ambiguities caused by occlusion, reflection, missing texture and illumination changes. In addition, the application scope of this dataset includes improving the accuracy and efficiency of optical flow estimation in computer vision, particularly when addressing optical flow-related issues in complex scenarios and dynamically changing environments.