Unsupervised Domain Adaptive Image Semantic Segmentation Based on Convolutional Fine-Grained Discriminant and Entropy Minimization

Abstract

AbstractDeep convolutional neural networks have made considerable progress in the field of semantic segmentation of images. However, due to inter-domain differences, even modern networks cannot segment test datasets from different domains very well. To reduce and avoid costly annotation of the source domain training data, unsupervised domain adaptation attempts to provide efficient information transfer from the source domain with detailed annotation to the target domain without annotation. However, most existing methods attempt to align the source and target domains from a holistic view, ignoring the underlying class-level structure in the target domain, along with large noise and ambiguity at the class junctions. In this work, we innovatively employ a fine-grained unsupervised domain adaptation semantic segmentation method with increased entropy certainty, and guide the model for finer-grained feature alignment by adversarial learning, while increasing the pixel certainty near the category boundaries. Our approach is easy to implement and we have achieved good results on both the urban road scene datasets GTA5->Cityscapes and SYNTHIA->Cityscapes.KeywordsSemantic SegmentationUnsupervised Domain AdaptationClass-Level Alignment