Exploring image captioning datasets
Several datasets are available for captioning image task. The datasets are usually prepared by showing an image to a few persons and asking them to write a sentence each about the image. Through this method, several captions are generated for the same image. Having multiple options of captions helps in better generalization. The difficulty lies in the ranking of model performance. For each generation, preferably, a human has to evaluate the caption. Automatic evaluation is difficult for this task. Let's explore the Flickr8
dataset.
Downloading the dataset
Flickr8
is gathered from Flickr and is not permitted for commercial usage. Download the Flickr8
dataset from https://forms.illinois.edu/sec/1713398. The descriptions can be found at http://nlp.cs.illinois.edu/HockenmaierGroup/8k-pictures.html. Download the text and images separately. Access to it can be obtained by filling in a form shown on the page:

An email will be sent with the download link. Once downloaded...