Instance Adaptive Self-Training for Unsupervised Domain Adaptation

Abstract

The divergence between labeled training data and unlabeled testing data is a significant challenge for recent deep learning models. Unsupervised domain adaptation (UDA) attempts to solve such a problem. Recent works show that self-training is a powerful approach to UDA. However, existing methods have difficulty in balancing scalability and performance. In this paper, we propose an instance adaptive self-training framework for UDA on the task of semantic segmentation. To effectively improve the quality of pseudo-labels, we develop a novel pseudo-label generation strategy with an instance adaptive selector. Besides, we propose the region-guided regularization to smooth the pseudo-label region and sharpen the non-pseudo-label region. Our method is so concise and efficient that it is easy to be generalized to other unsupervised domain adaptation methods. Experiments on 'GTA5 to Cityscapes' and 'SYN-THIA to Cityscapes' demonstrate the superior performance of our approach compared with the state-of-the-art methods.

Multiple Fusion Adaptation: A Strong Framework for Unsupervised Semantic Segmentation Adaptation

“…In MFA, the ensemble of two warmup models further benefits the accuracy of the pseudo labels. However, the offline manner is prone to the problems of ignorance of hard samples [10] and longer exposure to the potential noisy labels. We thus introduce the online pseudo labels as a supplementary.…”

“…Among the iterations of "assigning pseudo label" and "re-training", the pseudo labels are updated after several training epochs, yielding the "offline" manner. While the offline manner allows additional postprocessing and has the advantage of balanced pseudo labels [24], it is prone to the ignorance of hard and informative samples [10]. It is because the offline manner only preserves the most confident predictions among all the target domain data, which are relatively easy.…”

LabOR: Labeling Only if Required for Domain Adaptive Semantic Segmentation

“…Therefore, unsupervised domain adaptation (UDA) [15] addresses this problem by utilizing and transferring the knowledge of label-rich data (source data) to unlabeled data (target data), which can reduce the labeling cost dramatically [38]. According to the adaptation methodology, UDA can be largely divided into Adversarial learning based [27,49,50,52] DA and Self-training based [31,34,45,55,57] DA. While the former focuses on minimizing task-specific loss for Our novel human-in-the-loop framework, LabOR (PPL and SPL) significantly outperforms not only previous UDA state-of-the-art models (e.g., IAST [31]) but also DA model with few labels(e.g., WDA [36]).…”

“…According to the adaptation methodology, UDA can be largely divided into Adversarial learning based [27,49,50,52] DA and Self-training based [31,34,45,55,57] DA. While the former focuses on minimizing task-specific loss for Our novel human-in-the-loop framework, LabOR (PPL and SPL) significantly outperforms not only previous UDA state-of-the-art models (e.g., IAST [31]) but also DA model with few labels(e.g., WDA [36]). Note that our PPL requires negligible number of label to achieve such performance improvements (25 labeled points per image), and our SPL shows the performance comparable with fully supervised learning (0.1% mIoU gap).…”

“…source domain and domain adversarial loss, the self-training strategy retrains the model with generated target-specific pseudo labels. Among them, IAST [31] achieves state-ofthe-art performance in UDA by effectively mixing adversarial based and self-training based strategies.…”