Evidential Graph Contrastive Alignment for Source-Free Blending-Target Domain Adaptation

TL;DR

This work tackles Source-Free Blending-Target Domain Adaptation (SF-BTDA), where a single source model must be adapted to a mixed set of unlabeled target domains without source data. The authors introduce Evidential Contrastive Alignment (ECA), which combines Calibrated Evidential Learning to refine pseudo-label quality with a graph-based contrastive module guided by a Domain Distance Matrix to align samples of the same class across diverse target styles. The method jointly optimizes a calibrated evidential loss $L_{CEL}$ and a graph contrastive loss $L_{CON}^e$, achieving state-of-the-art results on Office-31, Office-Home, and DomainNet without domain labels or access to source data. Theoretical insights and extensive ablations support the effectiveness of both high-quality pseudo-label selection and cross-domain class alignment, with practical implications for real-world domain adaptation under privacy and labeling constraints.

Abstract

In this paper, we firstly tackle a more realistic Domain Adaptation (DA) setting: Source-Free Blending-Target Domain Adaptation (SF-BTDA), where we can not access to source domain data while facing mixed multiple target domains without any domain labels in prior. Compared to existing DA scenarios, SF-BTDA generally faces the co-existence of different label shifts in different targets, along with noisy target pseudo labels generated from the source model. In this paper, we propose a new method called Evidential Contrastive Alignment (ECA) to decouple the blending target domain and alleviate the effect from noisy target pseudo labels. First, to improve the quality of pseudo target labels, we propose a calibrated evidential learning module to iteratively improve both the accuracy and certainty of the resulting model and adaptively generate high-quality pseudo target labels. Second, we design a graph contrastive learning with the domain distance matrix and confidence-uncertainty criterion, to minimize the distribution gap of samples of a same class in the blended target domains, which alleviates the co-existence of different label shifts in blended targets. We conduct a new benchmark based on three standard DA datasets and ECA outperforms other methods with considerable gains and achieves comparable results compared with those that have domain labels or source data in prior.

Figures (4)

• Figure 1: Different DA settings. (a) Single Target Domain Adaptation (STDA). (b) Source-Free STDA (SF-STDA). (c) Multiple Target Domain Adaptation (MTDA). (d) Source-Free MTDA (SF-MTDA). (e) Blending-Target Domain Adaptation (BTDA). (f) Source-Free BTDA (SF-BTDA), a new DA setting proposed in this paper, with only access to source trained model.
• Figure 2: The framework for our proposed Evidential Contrastive Alignment (ECA). (a) Source Data Training: We train the model through single source domain and we are not access to the source domain data. (b) Domain Distance Calculation: We generate the mixture domain distance through low-level features and original image textures in an unsupervised way. (c) Calibrated Evidential Learning: We propose a calibrated evidential learning module to iteratively improve both the accuracy and certainty of the resulting model and adaptively generate high-quality pseudo target label (addressing Challenge 1). (d) Graph Contrastive Learning: We design a graph contrastive learning with the domain distance matrix and confidence-uncertainty criterion, to minimize the distribution gap of samples of a same class in the blended target domains, which alleviates the co-existence of different label shifts in blended targets (addressing Challenge 2).
• Figure 3: Overall Accuracy (%) on Office-31 and OfficeHome for ablation studies.
• Figure 4: Comparisons among other contrastive learning methods and other uncertainty methods with our ECA.