Supervisors: Christoph Mayer and Dr. Radu Timofte
n this thesis we tackle the problem of facial attractiveness regression and gender classification using deep neural networks. We use a large scale data set where the attractiveness was obtained from a mobile dating app. Instead of directly regressing the attractiveness scores, we transform it into a classification problem. We use the Person correlation to assess the quality of our models, because we are mainly interested in the relative attractiveness ordering between different faces, e.g. ranking a selection of images based on their attractiveness. We achieve a Pearson correlation of 0.635 on our test set. Furthermore, we evaluate our pretrained method on the SCUT-FBP5500 dataset, where we achieve state-of-the-art results. In a second step, we add the capability of gender classification to our network that then forms a multi-task network. When tackling both tasks within the same network the performance of our model drops only slightly compared to single-task networks but achieves a smaller run-time than requiring two individual models.