PVGCRA: Prediction Variance Guided Cross Region Domain Adaptation

Abstract

Semantic segmentation has become a very important task in computer vision in recent years, however, it usually requires a large amount of labeled data matching the considered scene to obtain reliable performance. Collecting and labeling large datasets for each new task and domain is very expensive, time-consuming and error-prone. To cope with this, unsupervised domain adaptation methods have been attempted for semantic segmentation tasks. Existing methods still suffer from poor class-level feature alignment and pseudo labels contain much noise. The latest method Cross Region Domain Adaptation (CRA) applies adversarial training to align the feature distribution of trusted and untrusted regions of the target domain image. In this paper, we re-model the uncertainty estimation module and propose to use the prediction variance as an uncertainty estimation method to align the feature distribution in the new trusted and untrusted regions. This approach is simply called Prediction Variance Guided Cross Region Domain Adaptation (PVGCRA). Experiments on the typical unsupervised domain-adaptive semantic segmentation task scenario GTA5 → Cityscapes show that this method improves the performance of the segmentation model and possesses better performance.