Few-shot domain adaptation for semantic segmentation

Unsupervised domain adaptation (UDA) has attracted increasing attention in recent years because it can liberate the labor force of annotating data. However, in many cases, the target data is scarce because it is difficult to obtain, at this time, the supervised domain adaptation (SDA) becomes attractive. In this work, we propose a novel few-shot supervised domain adaptation framework for semantic segmentation. The main idea is to exploit adversarial learning to align the features extracted from networks. To address the challenge of scarce data, we propose a pairing method of creating pairs using source data and target data. We design our framework as a two-stage structure to enhance the adaptation of the low-level features. In order to ensure the stable and effective training, we employ the spectral normalization in the discriminators and propose an alternately training strategy for the whole framework. Our proposed framework can work well even when there is only one sample per category. We evaluate our proposed method on a challenging synthetic dataset to real-world dataset adaptation where the results demonstrate the effectiveness of our method.