Large-scale vision-language pre-training has achieved promising results on
downstream tasks. Existing methods highly rely on the assumption that the
image-text pairs crawled from the Internet are in perfect one-to-one
correspondence. However, in real scenarios, this assumption can be d