MixedGAN: Facilitating GAN for Domain Adaptation Learning based on Mixing Different Domain Images for Semantic Segmentation

Yuehua Song, WonSook Lee

Abstract: Convolutional neural network-based semantic segmentation methods rely on the supervision of pixel-level labels, but since pixel-level labelling is time-consuming and labor-intensive, research has shifted to Unsupervised domain adap-tation (UDA), which is mainly used to migrate the learned knowledge from one domain to another, so that synthetic data can be used to help the model learn. Synthetic data can be used to help the model to learn, but the accuracy is not high compared to supervised learning. The reason for the low accuracy of UDA is the gap between domains. In this paper, a new UDA model is proposed to reduce the gap between domains in two steps. The first step is to mix images from two domains and generate pseudo-labels for the generated images, which helps to close the image distribution between domains. In the second step, the Generating Adversarial Neural Network (GAN) is used to distinguish the source of the feature map to further reduce the inter-domain gap. After several experiments, the accuracy of this model is much better than some classical DA models.