--- license: creativeml-openrail-m task_categories: - zero-shot-classification tags: - code size_categories: - 1K ComVG benchmark aims to test vision-language models ability in text-to-image retrieval.
We selected 542 high-quality images from Visual Genome and created 5400 datapoints in ComVG.
Each datapoint contains a postive and negative image. The negative image is a mutated variant with singular discrepancies in subject, object, or predicate.
For more details on creation process, please refer: https://arxiv.org/abs/2211.13854