Core Concepts
COCA method utilizes textual prototypes for few-shot learners in SF-UniDA, outperforming existing models.
Abstract
Universal domain adaptation (UniDA) addresses domain and category shifts.
Source-free UniDA (SF-UniDA) eliminates the need for direct access to source samples.
COCA method focuses on classifier optimization for SF-UniDA challenges.
Utilizes textual prototypes to distinguish common and unknown classes.
Experiments show COCA outperforms state-of-the-art models in UniDA and SF-UniDA.
Stats
COCA는 state-of-the-art UniDA 및 SF-UniDA 모델을 능가하는 성능을 보여줍니다.
Quotes
"COCA is a new paradigm to tackle SF-UniDA challenges based on VLMs."
"Experiments show that COCA outperforms state-of-the-art UniDA and SF-UniDA models."