Revisiting Unsupervised Domain Adaptation Models: a Smoothness Perspective

Xiaodong Wang, Junbao Zhuo, Mengru Zhang, Shuhui Wang, Yuejian Fang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1504-1521

Unsupervised Domain Adaptation (UDA) aims to leverage the labeled source data and unlabeled target data to generalize better in the target domain. UDA methods utilize better domain alignment or carefully-designed regularizations to increase the discriminability of target features. However, most methods focus on directly increasing the distance between cluster centers of target features, i.e., enlarging inter-class variance, which intuitively increases the discriminability of target features and is easy to implement. However, due to intra-class variance optimization being under-explored, there are still some samples of the same class are prone to be classified into several classes. To handle this problem, we aim to equip UDA methods with the high smoothness constraint. We first define the model's smoothness as the predictions similarity within each class, and propose a simple yet effective technique LeCo (impLicit smoothness Constraint) to promote the smoothness. We construct the weak and strong "views" of each target sample and enforce the model predictions of these two views to be consistent. Besides, a new uncertainty measure named Instance Class Confusion conditions the consistency is proposed to guarantee the transferability. LeCo implicitly reduces the model sensitivity to perturbations for target samples and guarantees smaller intra-class variance. Extensive experiments show that the proposed technique improves various baseline approaches by a large margin, and helps yield comparable results to the state-of-the-arts on four public datasets. Our codes are publicly available at https://github.com/Wang-Xiaodong1899/LeCo_UDA.