Shape robustness in style enhanced cross domain semantic segmentation. (March 2023)

• Record Type: Journal Article
• Authors: Zhu, Siyu; Tian, Yingjie
• Is Part Of: Pattern recognition. Volume 135(2023)
• Publication Date: 2023-03
• Subjects: Domain adaptation -- Semantic segmentation -- Transfer learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.109143
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We further improved the quantitative experiment including color augmentation technique. We carefully check the consistency of tense in related work and the normalization in reference list. To better present parameter learning, we replace the table with a curve figure. According to comments we further discussed some work related to semantic segmentation and transformer. Abstract: This paper focuses on domain adaptation method based on style transfer. Previous methods based on style transfer pay attention to the transformation of texture features between domains and maintain semantic consistency to the greatest extent. However, these methods have different effects on domain gaps in different types of categories. The categories with large texture difference and small structure difference can be improved better. For the categories with small texture difference and large structure difference, it causes negative transfer. In this paper, a shape robustness enhanced domain adaptive segmentation algorithm is proposed. Firstly, we adopt adjustable style transfer methods to enhance the style diversity of source domain images. Next, we differentiated different types of image features to weaken the negative transfer in the process of adversarial training. The results of this paper on general data sets GTA5 and SYNTHIA are better than other style transfer methods. Further experiments show that we improve the shape robustness of style enhancement method in domain adaptive segmentation task.