Conceptual Captions 12M

[Original Dataset] [Dataset Download]

Conceptual 12M (CC12M), a dataset with ~12 million image-text pairs meant to be used for vision-and-language pre-training. It is larger and covers a much more diverse set of visual concepts than the Conceptual Captions (CC3M), a dataset that is widely used for pre-training and end-to-end training of image captioning models. Check our paper for further details.

Citations

@inproceedings{changpinyo2021cc12m,
  title = {{Conceptual 12M}: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts},
  author = {Changpinyo, Soravit and Sharma, Piyush and Ding, Nan and Soricut, Radu},
  booktitle = {CVPR},
  year = {2021},
}