Master Assignment
Compositional scene generation using generative adversarial networks
Type: Master M-EE
Location: UvA
Period: Jan, 2018 - Jul, 2018
Student: Türkoğlu, M.Ö. (Ozgur, Student M-EE)
Date final project: July 16, 2018
Supervisors:
Abstract:
Image generation is an interesting problem in machine learning and computer vision. It has recently received a lot of attention due to the widespread applications based on generative adversarial models (GANs). These models can for example generate faces, animals, or simple scene in a realistic manner. However, the user cannot have a control over the scene layout.
Indeed, there is no control over the components occurring in the generated image. A mechanism is actually desirable to easily manipulate the elements that compose a scene. For example, it is currently hard to tell a model to generate "a bedroom with a nightstand and an armchair".
The objective of this master's thesis is to propose a methodology for compositional scene generation. The model will generate a scene in a sequential manner from pre-defined individual elements. Generating a scene step-by-step should provide explicit ,control over what and where elements are added in the scene in addition to more realistic, and diverse images compared to conventional GANs.