I’m excited to share that our paper “Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation” (led by Long Vuong) has been accepted to CVPR 2025! :fire::fire::fire:

While CLIP-based methods for Unsupervised Domain Adaptation (UDA) have shown promise, they face limitations in target domain generalization due to embedding distribution shifts. In this paper, we propose a novel approach that exploits the geometric relationships between visual and text embeddings through optimal transport theory. By leveraging clustering behavior in multi-modal embeddings and reference predictions from source prompts, our method achieves superior performance in target-prompt learning and representation quality.