Optasia/captioned_images

Dataset Summary This is a 660+ image dataset captioned professionally, part of a 450M image dataset with 780M records of ground truth. This is a highly diverse, interleaved dataset. Many images are highly aesthetic and many are everyday photos taken by tens of millions of people across 8 years with different cameras in different settings, captioned descriptively and accurately by hand. They were used to train ML Vision models.

PII and images of humans have been removed from this sample, captions are in the image_captions.csv file. You can request access to the entire dataset by emailing optasia.corp@gmail.com.

You can expect a wide variety of objects: 295601 gift cards, 12029267 shirts, 4497752 dogs, 1855440 trees, 13222 lug car wheels, 15201 geckos, 18294 wine bottle racks, etc. Furniture, humans, cars, trees, bugs, animals, buildings, etc are all captioned.

Languages The text in the dataset is mostly English. It also contains other languages such as Spanish, Arabic, etc.