Progressive Image generation between large gap domains with hierarchical disentangled representations

The School of EECS is hosting the following PhD progress review 2 seminar

Progressive Image generation between large gap domains with hierarchical disentangled representations

Speaker: Tianren Wang
Host: Prof Brian Lovell

Abstract: Image generation task aims to generate a desired image from any given input which can be either noise vector, original image or text description. In particular, translating an input image from the original domain to another domain is also referred as the image to image (I2I) translation task. The image generation and translation tasks have gain popularity as it can be used in a wide range of applications such as closing domain gap, photo manipulation, and synthesis training data. Recently, Generative Adversarial Network (GAN) has become an important and promising tool for solving image translation tasks. This project will utilize disentangled representation learning and GAN to translate samples from one domain to another as well as generating images from text description. Once disentangled, the representations are disentangled into several factors (i.e. face identity, gender, etc). Unfortunately, training the representation requires labels for each factor. When the labels are not available, the learning process resorts to the latent space assumption which defines the existence of a common shared latent space between two domains. Often this assumption is ineffective as there is no explicit definition of the latent space. Therefore, this project will study the explicit form of such a latent space to make the learning process more accurate and interpretable. Furthermore, we propose to disentangle the unnecessary feature entanglement while simultaneously keeping the useful entanglement by classifying the features hierarchically. In this way, we can keep the high-level content-related feature entanglements while disentangling the low-level feature entanglements.