Master Assignment
Compositional scene generation using generative adversarial networks
Type: Master M-EE
Location: University of Twente
Duration: Nov, 2018 - Jul, 2019
Student: Turkoglu, M.O. (Mehmet Ozgur, Student M-EE)
Final project: July 15, 2019
Supervisors:
Abstract:
In this master’s thesis, a novel sequential image generation model based on Generative Adversarial Networks (GANs) is proposed. Even though recent GAN-based approaches have been successful in generating for example faces, birds, flowers, street view images in a realistic manner, user control over the image is still limited. The proposed approach generates an image element-by-element (object-by-object) progressively and improves the controllability of the image generation process explicitly through an element-specific latent vector. Also, it improves the controllability by resolving affine transformation and occlusion issues existing conditional GANs models have. Experiments are carried out on the subset of the challenging and diverse MS-COCO dataset and the proposed model is compared with the state-of-the-art baselines. Both qualitative and quantitative results are provided to show the strength and the advantages of the proposed model.