(i)

Colophon



The central goal of computer vision systems is to make sense of visual material. This is achieved by feeding large-scale collections of semantically enriched images – where each element in each frame has been explicitly described by a human being – to artificial agents. Based on this ‘reading’ of the source material, Artificial Intelligence learns to discern what our world looks like. Most of the time such datasets are its only teachers, and thus the datasets play a crucial role in machine learning. When studying them carefully, it becomes clear that in general little attention is paid to the quality and the origins of the data which carry – together with information – also unfairness and biases.

The current website displays all the elements from a specific dataset: MS COCO. Created by Microsoft Research, COCO shows many examples of common objects in context and offers ground truth guidelines for many computer vision systems. COCO has been created by web scraping a large collection of images from Flickr. As a consequence, this gives an implicit US-centred connotation to the resource.

This website is an attempt to reverse engineering the resource in order to geolocate all of its original images and link them to their original Flickr links. In the form of an interactive database, the website wants to emphasise the geographical biases embedded within the systems.

Sorry, go to your desktop browser.