Training-free label space alignment for universal domain adaptation

초록

Universal domain adaptation (UniDA) transfers knowledge from a labeled source domain to an unlabeled target domain, where label spaces often differ and the target domain often contains private classes. Previous UniDA methods primarily focused on visual space alignment, however, severe domain shifts often lead to visual ambiguities, undermining their ability to predict the label space shift. Our preliminary experiment shows that focusing exclusively on label space alignment can be a more effective solution to these limitations. While generative vision-language models (VLMs) offer strong zero-shot capabilities that could help uncover target-private classes, naively employing them introduces challenges such as noisy and semantically ambiguous labels. To address these challenges, we propose a novel approach that carefully utilizes generative VLMs to improve label space alignment. We introduce an adaptive thresholding strategy that leverages meaningful relationships between source and discovered target labels. This allows a training-free mechanism to filter out noisy and ambiguous labels, effectively refining the label space by identifying shared and discovering target-private label subsets. Building upon this aligned space, we subsequently employ a robust self-training stage to maximize discriminability using the reliable pseudo-labels derived from the identified label spaces. The results reveal that the proposed method considerably outperforms existing UniDA techniques across key benchmarks, delivering an average improvement of +7.9% in H-score and +6.1% in H3-score. Furthermore, incorporating self-training further enhances performance, yielding additional increments of +1.6% in the H-score and +2.7% in the H3-score.

키워드