Open X-funbean opened 1 week ago
It is mentioned in the analysis of the paper that the semantic space of the image is limited, so whether more negative labels will "just" cover the OOD data in the test set, resulting in the improvement of zero-shot performance?
You can refer to our rebuttals in ICLR 2024 and a reviewer asked a similar question.
It is mentioned in the analysis of the paper that the semantic space of the image is limited, so whether more negative labels will "just" cover the OOD data in the test set, resulting in the improvement of zero-shot performance?