Machine learning for the comparison of synthetic images as a tool to support the evaluation of rendering exams

The aim of this project is to design and develop a system to compare two similar synthetic images and identify and classify their differences. These images, referred to as Reference and Render, are generated with Blender’s Cycles, a ray-trace-based production render engine capable of ultra- realistic rendering. The Reference image and the corresponding 3D meshes are provided to bachelor’s students, who should prove their capabilities in rendering for design by reproducing all the visible features of the Reference from the same viewpoint, generating a new Render image. The system should support both students and teachers in identifying and explaining the differences between the Reference and Render image. The problem has been addressed with a machine learning approach: to perform the comparison, a neural network for semantic change detection was trained using a newly annotated dataset generated with Blender. This dataset enabled the network to distinguish both the target features, such as textures, shadows and transparencies and the changes related to them. The model performance is evaluated with metrics as precision, recall and F-score, whereas the extent of the differences has been evaluated by means of similarity measures.