Adapt via Bayesian Nonparametric Clustering: Fine-Grained Classification for Model Recycling Under Domain and Category Shift

Recycling pretrained classification models for new domains, known as Source-Free Domain Adaptation (SFDA), has been extensively studied under the closed-set assumption that source and target domains share identical label spaces. However, this assumption does not hold when unseen classes appear in the target domain. Addressing this category shift is challenging, as unknown target classes usually arise with no prior knowledge of their identities or number, and becomes particularly difficult in the source-free setting, where access to source data is unavailable. Most existing methods treat all unknown classes as a single group during both training and evaluation, limiting their capacity to model the underlying structure within the unknown class space. In this work, we present Adapt via Bayesian Nonparametric Clustering (ABC), a novel framework designed for SFDA scenarios where unknown target classes are present. Unlike prior methods, ABC explicitly achieves fine-grained classification of unknown target classes, offering a more structured vision of the problem. Our method first identifies high-confidence target samples likely to belong to known source classes. Using these as guidance, we develop a guided Bayesian nonparametric clustering approach that learns distinct prototypes for both known and unknown classes without requiring the number of unknown classes a priori, and assigns target samples accordingly. We further introduce a training objective that refines the source model by encouraging prototype-based discriminability and local prediction consistency. Experiments show that our method achieves competitive performance on standard benchmarks while simultaneously providing effective clustering of unknown classes.