基于合成图像的语义分割任务域适应算法研究

Research on Domain Adaptation Algorithm for Semantic Segmentation Task Based on Synthetic Image

深度域适应是计算机视觉中极为重要的课题,对于获取手工标记数据非常困难和繁琐的场景比如语义分割任务,解决域适应问题尤为重要。先前的研究表明,即使是深度神经网络也无法很好地学习跨域的信息表示。论文专注于在语义分割场景下,调整分割网络在源域(合成图像)和目标域(真实图像)中学习到的特征表示。与之前使用简单对抗性目标或超像素信息来辅助的方法不同,论文提出了一种基于生成对抗网络(GAN)的方法,该方法使不同域中的特征表示在学习到的特征空间中更接近。实验结果表明论文提出的方法可以在合成图像域到真实图像域的一个具有挑战性的场景中实现较为先进的结果。

Deep domain adaptation is an extremely important topic in computer vision,and it is especially important to solve the domain adaptation problem in scenarios where it is very difficult and cumbersome to obtain manual labeled data,such as semantic segmentation tasks.Previous research has shown that even deep neural networks do not learn well the representation of information across domains.This paper focuses on adjusting the feature representations learned by the segmentation network in the source domain(synthetic image)and target domain(real image)under the semantic segmentation scenario.Different from previous methods that use simple adversarial targets or superpixel information to assist,this paper proposes a generative adversarial network(GAN)-based method that brings feature representations in different domains closer in the learned feature space.Experimental results show that the proposed method can achieve state-of-the-art results in a challenging scenario from the synthetic image domain to the real image domain.