A Unified Approach to Universal Domain Adaptation with Single and Multiple Source Domains

Junsong Leng, Zhong Chen, Guoyou Wang

Keywords: Universal Domain Adaptation, Multi-modal, Uncertainty Estimation

Abstract: Universal domain adaptation (UniDA) imposes no constraints on the label sets of the source and target domains, aiming to transfer knowledge from source to target domains. Existing works typically target either single-source or multi-source UniDA, rarely both. Naively merging multi-source data into a single source domain may lead to negative transfer and performance degradation. Moreover, since multi-source models are often equipped with modules tailored for multi-source data, they are usually not directly applicable to single-source tasks. These challenges hinder the development of a unified framework. In this paper, we propose a unified model based on multi-modal and uncertainty estimation, termed MUEUDA, to address this issue. Our model is capable of effectively handling both single-source and multi-source settings with outstanding performance. First, we incorporate multi-modal information, enabling class-level feature alignment between source and target domains using fine-tuning and prompt learning techniques. Second, we extract class-level image feature prototype from the source domain and progressively update them during training. Finally, we introduce a novel uncertainty estimation method that determines whether an image in the target domain belongs to a known or unknown class through a learnable threshold. Extensive experiments are conducted on both single-source and multi-source benchmarks, and our model achieved state-of-the-art performance. The method demonstrates strong performance across both scenarios, balancing effectiveness and generality. The code is available at https://github.com/jstree365/MUEUDA.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 7650