基于位置对抗学习的道路场景无监督域自适应语义分割

Unsupervised Domain Adaptation Semantic Segmentation of Road Scene Based on Location Adversarial Learning

在城市道路场景图像的无监督域自适应语义分割任务中,现有的基于对抗学习的跨域语义分割方法往往会忽视图像中的空间位置关系,在图像整体层面上进行对抗,这会导致卷积神经网络偏向于提取两个域之间主要类别的特征。道路场景图像中的上、中、下3部分对应的类别不同且类别占比差距也较大,但以往的方法没有充分利用图像的空间位置结构,也没有考虑到数据集中存在的类别占比数量不平衡的问题。为了解决这个问题,提出了横向位置分块对抗和纵向位置分块对抗的方法。在块与块内做对抗损失,这样可以使域之间类别的靠拢更加细节化,一定程度上解决了数据集中的类别数量不平衡问题。通过在数据集GTA5到Cityscapes和SYNTHIA到Cityscapes上的实验,证明了所提方法的有效性。

In the unsupervised domain adaptive semantic segmentation task of urban road scene images,the existing cross-domain semantic segmentation methods based on adversarial learning tend to ignore the spatial positional relationship in the image and conduct confrontation at the overall level of the image,which will lead to convoluted The product neural network is biased towards extracting the features of the main categories between the two domains.The upper,middle,and lower parts of the road scene image correspond to different categories and the difference in category proportions is also large.However,the previous methods did not make full use of the spatial position structure of the image,and did not take into account the number of categories in the data set.imbalance problem.In order to solve this problem,horizontal position block adversarial and vertical position block adversarial methods are proposed.Do adversarial loss between blocks,which can make the closeness of categories between domains more detailed,and solve the problem of the imbalance of the number of categories in the dataset to a certain extent.The effectiveness of the proposed method is demonstrated by experiments on datasets GTA5 to Cityscapes and SYNTHIA to Cityscapes.