Algorithmic Ways of Seeing: Improving Image Recognition by Training on Art Images

Image recognition systems perform well when detecting objects in photographic images, but poorly when confronted with an image style that is different from the data used in training - in particular when dealing with non-photographic images. This is remarkably different from human vision: Even a very small child can recognize the same objects in photographies and simple drawings. We hypothesize that developing visual recognition algorithms that can recognize depictions in art images may give us a deeper understanding of computer vision compared to human vision, and may help develop ​more general and robust vision algorithms.​ We address two important obstacles: lack of sufficiently large datasets of non-photographic images for training, as well as the ethical and cultural tensions between computer vision and human interpretation of images.