Video Semantic Segmentation for Intersection by Domain Adaptation

In recent years, semantic segmentation, which is one of the image recognition technologies for automatic driving, has attracted attention. Semantic segmentation can perform high accuracy object detection by discriminating classes with pixel level precision. However, training data for segmentation usually requires extensive manual labor because a corresponding label has to be assigned to each pixel. Currently, using a computer graphics (CG) dataset makes it easier to create supervised data. However, when inference images are real-world, intersection over union (IoU) is greatly reduced by the variation of the domain. In this work, an unsupervised domain adaptation (UDA) training method is proposed to achieve efficient intersection video segmentation models for each target domain. The inference is performed for intersection images with relatively complex situations. The simulation results show that the proposed algorithm for semantic segmentation of intersection videos achieve IoU scores comparable to supervised learning for classes that are prominently displayed on the screen.