Abstract:

The adaptive segmentation network represents an efficacious method for cross-domain semantic segmentation within urban scenes. However, the challenge arises from the distinct appearance distributions among cross-domain datasets, leading to domain gaps and unsatisfactory network segmentation accuracy for small targets. To address these issues, a domain adaptive segmentation method based on a dual-source discriminator was proposed. Firstly, the new source domain S' was obtained using the style translation technology FastPhotoStyle for the source domain S, thereby reducing the domain gaps at the image level. Next, the generator was employed to extract segmentation feature maps from the source domain S, the new source domain S', and the target domain T, respectively. The feature map of the new source domain served as an intermediate bridge for the channel-wise fusion between the source and target domains feature maps. The two fused feature maps were input into the dual-source discriminator, with both the dual-source discriminator and the generator undergoing iterative training. Since the discriminator input of the proposed model consists of dual-source features, it is referred to as a dual-source discriminator. The two features from the dual-source input contained similar feature information, which further reduced domain differences at the feature level. To enhance segmentation accuracy, a self-training pseudo-label was introduced. At the same time, to address class imbalance issues during training, a class balance factor was incorporated into the loss function of the target domain, thereby enhancing the network’s ability to segment small targets. Experiments on two segmentation tasks GTA5→Cityscapes and SYNTHIA→Cityscapes demonstrated the advancement and effectiveness of the proposed method.

Key words: dual-source discriminator, adversarial learning, domain adaptation, semantic segmentation, self-training