A Two-Stage Generative Adversarial Approach for Domain Adaptive Semantic Segmentation

The abundance of light and the wide range of object colors in the image datasets gathered in daylight situations make it simpler for semantic segmentation networks to extract useful image information. Nevertheless, the existing model trained in the daytime setting struggles to accurately discriminate distinct objects in the nighttime scenes. To solve this problem, we suggest a two-stage generative adversarial network-based unsupervised domain adaptive semantic segmentation algorithm. In the first stage, a circular generative adversarial network (RoundGAN) based on cycle consistency is proposed to transform the annotated daytime images in the source domain into a style similar to that of nighttime images in the target domain. A new semantic segmentation network training architecture (SDNet) is proposed in the second stage, which is based on the concept of an adversarial network. In this architecture, the conventional semantic segmentation network is used as a sub-module of the training network, and an adversarial loss function is added. A combination of fully supervised and unsupervised training techniques is used to train the segmentation network. Experiments show that the final trained model can effectively segment classes in nighttime images.