The big picture and why it needs history

Computer vision is not great at context – historians are

ANU historian Marnie Hughes Warrington considers what networked cameras see and what they miss in her new essay on computer vision and how to correct for the misunderstanding images establish. It is a  new note for what will surely be a vast study of the uncharted ground where historical method and AI intersect.

She starts with vision of a regency gent, frock-coated, stocked and straw top hatted, who is not what he appears. We know that – photographs of Beau Brummell are light-on but computer-vision just presents what it is there

This matters, because image only makes meaning in context, and vision does not explain what it shows. “Saying that computer vision is bad history decision making does not will it away. If anything, we have an ethical imperative to improve it because humans make bad historical decisions too, and because computer vision initiatives can get stuck in human support limbo,” MHW argues. And the more images collected over more time the harder it becomes to work out what is recorded really reveals.

So how to improve the way data-sets of images are assembled? She’s glad you asked that. “Historians don’t work with simple sequences of yes or no decisions, and they write with hard questions about information imbalance and fairness in mind. The way they work is ‘pre-industrial’, and I mean that in the nicest possible way. They turn out different decisions when working with the same materials, and that variety can be important for helping computer scientists to think of more, and better ways of building decision networks.”


Subscribe

to get daily updates on what's happening in the world of Australian Higher Education