Weakly Supervised Domain Adaptation using Super-pixel labeling for Semantic Segmentation

Abstract

Deep learning for semantic segmentation requires a large amount of labeled data, but manually annotating images are very expensive and time consuming. To overcome the limitation, unsupervised domain adaptation methods adapt a segmentation model trained on a labeled source domain (synthetic data) to an unlabeled target domain (real-world scenes). However, the unsupervised methods have a poor performance than the supervised methods with target domain labels. In this paper, we propose a novel weakly supervised domain adaptation using super-pixel labeling for semantic segmentation. The proposed method reduces annotation cost by estimating a suitable labeling area calculated from the Entropy-based cost of a previously learned segmentation model. In addition, we generate the new pseudo-labels by applying fully connected Conditional Random Field model over the pseudo-labels obtained using an unsupervised domain adaptation. We show that our proposed method is a powerful approach for reducing annotation cost.