COCA: Classifier-Oriented Calibration for Source-Free Universal Domain Adaptation

The author presents a novel COCA method that leverages textual prototypes to address the challenges of SF-UniDA, focusing on classifier optimization instead of image encoder optimization.