A Two-stage Cascading Method Based on Finetuning in Semi-supervised Domain Adaptation Semantic Segmentation

Abstract

The traditional unsupervised domain adaptation (UDA) has achieved great success in many computer vision tasks, especially semantic segmentation, which requires high cost of pixel-wise annotations. However, the final performance of UDA method is still far behind that of supervised learning due to the lack of annotations. Researchers introduce the semi-supervised learning (SSL) and propose a more practical setting, semi-supervised domain adaptation (SSDA), that is, having labeled source domain data and a small number of labeled target domain data. To address the inter-domain gap, current researches focus on domain alignment by mixing annotated data from two domains, but we argue that adapting the target domain data distribution through model transfer is a better solution. In this paper, we propose a two-stage SSDA framework based on this assumption. Firstly, we adapt the model from the source domain to the labeled dataset in the target domain. To verify the assumption, we choose a basic transfer mode: finetuning. Then, to align the labeled subspace and the unlabeled subspace of the target domain, we choose teacher-student model with class-level data augmentation as the basis to realize online self-training. We also provide a deformation to solve overfitting on the target domain with a small number of annotated data. Extensive experiments on two synthetic-to-real benchmarks have demonstrated the correctness of our idea and the effectiveness of our method. In most SSDA scenarios, our approach can achieve supervised performance or even better.