Abstract: It is always well believed that pre-trained vision-language foundation models (e.g., CLIP) would substantially facilitate vision-language tasks. Nevertheless, there has been less evidence in ...