Abstract
In computer vision, researcher focus on developing better algorithms in order to predict human preference of images. In contrast, in experimental aesthetics, researchers try to understand the basics of aesthetic perception. Deep Learning has recently pervaded computer vision, at the cost that these black-box approaches are not easy to understand. While computers get better and better at discriminating between images of high and low aesthetic appeal, it is hard to draw any conclusions from the decision boundaries that these deep models draw. In our present work, we studied variances of deep features and ask whether we can find a difference in the activation patterns between art and non-art images. We demonstrate that we can discriminate between the two classes of images by using only two different types of variances defined on the first layer features with an accuracy of 93%. In interpreting these measures, we can distill two main properties (here called richness and variability) that characterize traditional artworks when combined. We will also report on a study that uses statistical image properties to demonstrate objective differences between traditional art, so-called Bad Art, and abstract art.
Prof. Dr. med. Dr. rer. nat. habil. Christoph Redies
Ort: INB Seminarraum - 12:30 Uhr
