Cross-Domain Adaptive Clustering for Semi-Supervised Domain Adaptation

Abstract

In semi-supervised domain adaptation, a few labeled samples per class in the target domain guide features of the remaining target samples to aggregate around them. However, the trained model cannot produce a highly discriminative feature representation for the target domain because the training data is dominated by labeled samples from the source domain. This could lead to disconnection between the labeled and unlabeled target samples as well as misalignment between unlabeled target samples and the source domain. In this paper, we propose a novel approach called Cross-domain Adaptive Clustering to address this problem. To achieve both inter-domain and intra-domain adaptation, we first introduce an adversarial adaptive clustering loss to group features of unlabeled target data into clusters and perform cluster-wise feature alignment across the source and target domains. We further apply pseudo labeling to unlabeled samples in the target domain and retain pseudolabels with high confidence. Pseudo labeling expands the number of "labeled" samples in each class in the target domain, and thus produces a more robust and powerful cluster core for each class to facilitate adversarial learning. Extensive experiments on benchmark datasets, including DomainNet, Office-Home and Office, demonstrate that our proposed approach achieves the state-of-the-art performance in semi-supervised domain adaptation.

Low-confidence Samples Matter for Domain Adaptation

“…Inspired by previous works [32,88,93] in domain adaptation, we add FixMatch [63] to the existing method to construct a stronger baseline. Specifically, let T (x) and T (x) denote the weakly and strongly augmented views for x ∈ D tu , respectively.…”

“…[83] breaks down the SSDA problem into two sub-problems, namely, semi-supervised problem and UDA problem, and then proposes to learn consistent predictions using co-training. CDAC [32] proposes an adversarial adaptive clustering loss to group features of unlabeled target data into clusters and perform cluster-wise feature alignment across domains. CLDA [62] employs class-wise contrastive learning to reduce the inter-domain gap and instance level contrastive alignment to minimize the intra-domain discrepancy.…”

“…ECACL [33] propose a holistic framework that incorporates multiple complementary domain alignment techniques. Among these methods, [32,33,44] adopt Fix-Match [63] or its variants. In this work, we also use Fix-Match to build a strong baseline, and add a regularization term [97] to get more reliable pseudo labels.…”

PointMatch: A Consistency Training Framework for Weakly Supervised Semantic Segmentation of 3D Point Clouds

“…Thus it combines the advantages of both consistency regularization and pseudo-labeling (or self-training). This approach has been applied in many domains, such as semi-supervised learning (SSL) [2,3,36,50], unsupervised learning (USL) [8,12], unsupervised domain adaptation (UDA) [6,35], and semi-supervised domain adaptation (SSDA) [21,22], all of which prove the effectiveness of consistency training in learning high-quality representations from label-scarce data. More recently, there are some works extending consistency training into other tasks, such as unsupervised domain adaptation for image segmentation [27] and semi-supervised 3D object detection [43].…”

“…Recently, consistency training is acknowledged as a powerful algorithmic paradigm for robust learning from label-scarce data, e.g. in unsupervised/semi-supervised learning [8,12,36,50] and unsupervised/semi-supervised domain adaptation [6,21,22,35]. It works by forcing the model to make consistent prediction under different perturbations/augmentations to the input sample (named as different views) and the prediction in one view usually serve as the pseudo-label of the other view.…”