Computational aesthetics using machine learning for video game camera direction

Erdem, Ali Naci
Computational aesthetics is a developing field which employs computational approaches either for generating or evaluating aesthetic values. In the scope of this thesis, visual aesthetic quality of computer generated images was aimed to be improved using a computational aesthetics approach. An appropriate machine learning algorithm was selected and trained on a set of reference images collected online. Using the trained model, a novel video game camera direction method predicting the aesthetic quality of the real-time graphics and changing the virtual camera position accordingly was developed. In order for the proposed approach to be effective, a regression analysis assigning aesthetic quality values to images was utilized instead of high and low quality classification. Rather than dealing with semantic context, color distribution and compositional properties affecting aesthetic appeal were preferred and to make quicker aesthetic score predictions, faster and more efficient features were selected, considering their aesthetic foundations. Some of the existing features were improved, and some were tailored to be applied to regression analysis. Aesthetics being a highly subjective topic, only outdoor scene and landscape visuals were targeted in this work in order to reduce complexity. The proposed method on the other hand, can be extended to other environments by changing the training data. The prediction performance of the machine learning model was not very significant when compared to the previous works, yet promising considering the challenges and limitations involved and showed that a near-real time aesthetic analysis and visual improvement was possible through a “virtual” camera director.
Citation Formats
A. N. Erdem, “Computational aesthetics using machine learning for video game camera direction,” M.S. - Master of Science, Middle East Technical University, 2015.