Reproduce “Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation” on OpenResearch

About this paper

Weinan He, Zilei Wang, Yixin Zhang

Target Semantics Clustering (TASC) via Text Representations introduces a Universal Domain Adaptation (UniDA) framework that defines semantic centers within a discrete text representation space, leveraging Vision-Language Models to mitigate domain bias and adaptively estimate target clusters. The approach achieves state-of-the-art performance on four benchmarks, including substantial improvements of 6.0% on DomainNet and 11.3% on VisDA, across various category shift scenarios.

Key insights

• Defining semantic centers within a discrete text representation space, rather than unconstrained image feature space, intrinsically mitigates domain bias and provides appropriate semantic granularity for UniDA.
• Vision-Language Models can be deeply integrated into domain adaptation, moving beyond mere feature extraction to leverage their cross-modal semantic alignment capabilities for structuring the learning process.
• A two-stage optimization strategy can effectively handle the mixed-integer nature of discovering discrete semantic centers and refining continuous model parameters, simplifying complex UniDA objectives.
• It is possible to adaptively estimate the number of target clusters and detect unknown samples without prior knowledge, addressing key limitations of previous methods for robust universal adaptation.

Summary by alphaXiv.