D3GU: Multi-Target Active Domain Adaptation via Enhancing Domain Alignment

TL;DR

This work tackles multi-target active domain adaptation (MT-ADA) for image classification by introducing D$^3$GU, a unified framework that combines decomposed domain discrimination ($D^3$) to separately align source-target and target-target domains with a tunable balance $\alpha$, and Gradient Utility (GU) for joint, gradient-driven active sampling. GU is further paired with KMeans clustering (GU-KMeans) to ensure diversity while prioritizing samples that most influence both classification and domain alignment. Empirical results on Office31, OfficeHome, and DomainNet demonstrate state-of-the-art MT-ADA performance with a single model at inference and test-time domain-label freedom. The proposed approach advances practical MT-ADA by addressing both training-time alignment and budget-aware sample annotation, offering robust gains across diverse domain shifts.

Abstract

Unsupervised domain adaptation (UDA) for image classification has made remarkable progress in transferring classification knowledge from a labeled source domain to an unlabeled target domain, thanks to effective domain alignment techniques. Recently, in order to further improve performance on a target domain, many Single-Target Active Domain Adaptation (ST-ADA) methods have been proposed to identify and annotate the salient and exemplar target samples. However, it requires one model to be trained and deployed for each target domain and the domain label associated with each test sample. This largely restricts its application in the ubiquitous scenarios with multiple target domains. Therefore, we propose a Multi-Target Active Domain Adaptation (MT-ADA) framework for image classification, named D3GU, to simultaneously align different domains and actively select samples from them for annotation. This is the first research effort in this field to our best knowledge. D3GU applies Decomposed Domain Discrimination (D3) during training to achieve both source-target and target-target domain alignments. Then during active sampling, a Gradient Utility (GU) score is designed to weight every unlabeled target image by its contribution towards classification and domain alignment tasks, and is further combined with KMeans clustering to form GU-KMeans for diverse image sampling. Extensive experiments on three benchmark datasets, Office31, OfficeHome, and DomainNet, have been conducted to validate consistently superior performance of D3GU for MT-ADA.

Figures (6)

• Figure 1: Illustration of MT-ADA task on OfficeHome with source domain (art), three target domains (clipart, product, real), and five classification classes (alarm_clock, backpack, batteries, bed, bike). The pretraining stage trains a model $\theta_0$ through unsupervised domain adaptation on labeled source images $L_{\mathcal{S}}$ and unlabeled target images $\bigcup\{U_{\mathcal{T}_i}\}_{i=1}^3$. Multiple active learning stages then follow to iterate between active sampling and domain adapted training. At the $j$-th active learning stage, the active sampling function selects and annotates some unlabeled target images to expand the labeled target set $\bigcup\{L_{\mathcal{T}_i}\}_{i=1}^3$ using the model $\theta_{j-1}$ trained at the previous stage. The domain adapted training step then finetunes $\theta_{j-1}$ to obtain $\theta_{j}$ using all images and annotations. Model performance is evaluated on target domains at test time. Section \ref{['section:preliminary']} describes more details.
• Figure 2: Each domain has its transferable positive knowledge ("+"), and negative knowledge ("-") that is domain-specific and cannot be shared. Binary domain discrimination can lead to under-alignment between positives in the target domains while all-way domain discrimination leads to over-alignment of the negatives. Decomposed domain discrimination instead balances the two to enhance domain alignment.
• Figure 3: Gradient utility $\varphi$ measures each target sample's contributions towards classification ($\varphi_{cls}$) and domain alignment ($\varphi_{da}$). Target samples with the highest $\varphi$, rather than the most uncertain ones, are upweighted the most.
• Figure 4: Plot of MT-ADA accuracies at $4$ active learning stages. We plot classification accuracies averaged over all source domains on each dataset (AVG in Tables \ref{['tab:office31-home-exp']}, \ref{['tab:domainnet-exp']}).
• Figure 5: Top: UDA pretrain performances of different $\alpha$ values in Equation \ref{['equation:decompose-loss']} with quickdraw as source domain on DomainNet. Bottom: Effect of $\beta$ values in Equation \ref{['equation:beta-norm']} on GU-KMeans $+$ binary domain discrimination, with art as source domain on OfficeHome.