The software performs the task of matching photos of a person to similar looking people in paintings, in order to retrieve paintings given a query photo. This is challenging as paintings span many media (oil, ink, watercolour) and can vary tremendously in style (caricature, pop art, minimalist) and other elements such as angle, lighting and size may vary.
Detecting similarities between a photo and a painting may be trivial for human eyes, but poses a significant challenge for a computer.
The software utilises CNN face representation and discriminative learning to provide increased performance over alternative methods. It demonstrates that facial features generated using Convolutional Neural Networks (CNN) produced from a network learnt entirely on photos are able to generalise remarkably well to paintings of many different styles. Furthermore, the similarity between these features can be used to find photos and paintings of people that look eerily similar.
For additional background material and performance metrics please refer to the academic publication relating to this software: