Dataset

We have chosen to split the dataset in two gzipped files. You can download the images from Wikipedia by using the "Images" links above and all the textual part by clicking on the "Text" links.

First, the dataset consists of:

• images from Wikipedia;
• captions created by humans in the English language;
• captions created by humans in at least one of the following languages: Portuguese, Spanish, French or German;
• automatic translations of these captions, from English into at least one of the four languages mentioned above;
• the URL of an ImageNet image obtained by means of the application of a bag-of-visual-words method (on Wikipedia's and ImageNet's images), selecting the closest image in ImageNet to each of the Wikipedia's images.

The textual part of this dataset consists of the following files:

metadata.final.file-found

local path to the Wikipedia image

reference.final.file-found

Wikipedia image caption in the reference language

source.final.file-found

Wikipedia image caption in the source language, i.e. English

automatic-translation.final.file-found

Wikipedia image caption in the reference language, obtained by machine translation of the source caption

imagenet.final.file-found

path to the ImageNet image entry online

The five files described above contain information for one sentence (image caption) per line.