Cluster Alignment With a Teacher for Unsupervised Domain Adaptation

Abstract

Deep learning methods have shown promise in unsupervised domain adaptation, which aims to leverage a labeled source domain to learn a classifier for the unlabeled target domain with a different distribution. However, such methods typically learn a domain-invariant representation space to match the marginal distributions of the source and target domains, while ignoring their fine-level structures. In this paper, we propose Cluster Alignment with a Teacher (CAT) for unsupervised domain adaptation, which can effectively incorporate the discriminative clustering structures in both domains for better adaptation. Technically, CAT leverages an implicit ensembling teacher model to reliably discover the class-conditional structure in the feature space for the unlabeled target domain. Then CAT forces the features of both the source and the target domains to form discriminative class-conditional clusters and aligns the corresponding clusters across domains. Empirical results demonstrate that CAT achieves state-of-the-art results in several unsupervised domain adaptation scenarios.

Unsupervised Domain Adaptation via Structurally Regularized Deep Clustering

“…We empirically observe that for z t with arg max k p t k = k or z s with its true label y s = k, they are learned to drift together with its cluster centroid c k in the feature space, rather than to collapse to c k . In some cases, the class-wise distances between the source and target domains are even getting larger by applying (8) and (10), suggesting that the SRGenC objective (12) is indeed modulating the feature space learning via generative clustering, which is in contrast to existing methods [4], [5], [6], [8], [38], [48] that explicitly align the features across the two domains. Results of these empirical studies are presented in Fig.…”

Unsupervised Domain Adaptation in Semantic Segmentation: A Review

“…To overcome the lack of target semantic supervision, other approaches [67,16] resort to pseudo-labels directly from network predictions to discover target class-conditional structures in the feature space. Those structures are then exploited to perform a within-domain feature clusterization [16] and cross-domain feature alignment by centroid matching [67,16]. Starting from analogous premises, we extend a similar form of inter and intra class adaptation to the semantic segmentation scenario, by introducing additional modules that help to address the inherent increased complexity.…”

Learning Target-Domain-Specific Classifier for Partial Domain Adaptation

“…In addition, Xu et al [15] proposed to adapt feature norms of both domains to achieve equilibrium, which is free from relationship of the label spaces. Recently, some methods apply ensemble learning to improve the discrimination ability of extracted features [37], [38].…”

“…French et al [37] used a mean teacher model [39] to mine the target domain knowledge. Deng et al [38] explored the classconditional structure of the target domain with an ensemble teacher model.…”