Learning Geometry Compatibility with 3D Convolutional Neural Networks

Abstract

Modern video games offer substantial amounts of customization options. Manually testing the visual compatibility of all options is time-consuming and error-prone. Together with Ghost Games, we present a method of learning the visual compatibility between pairs of geometries. We introduce a transformation pipeline and model architecture, which we train on hand-labeled data. Furthermore, we explore a part of the hyperparameter space of our proposed architecture and extend it to accommodate confidence predictions. Finally, we run a quantitative study on the trained model and suggest improvements and extensions for future work.