Public defence in Computer Science, M.Res Yuxin Hou
Title of the doctoral thesis: Learning Latent Image Representations with Prior Knowledge
Opponent: Assistant Professor Oisin Mac Aodha, University of Edinburgh, Skotland
Custos: Assistant Professor Juho Kannala, Aalto University School of Science, Department of Computer Science
The public defence will be organised on campus.
The thesis is publicly displayed 10 days before the defence in the publication archive Aaltodoc of Aalto University.
Public defence announcement:
Computer vision is essential for developing artificial intelligence systems as it enables machines to understand the observed environment. It aims to extract information to describe the world from image data. In recent years, deep learning has become the dominant tool for computer vision applications because of the ability to learn low-dimensional latent representations from images by encoding the image through a number of non-linear layers. Though deep learning shows the promising potential of learning powerful latent image representations, it is not good for interpretability and controllability. On the other hand, since we have prior knowledge for many applications already, we can utilize it to learn better latent image representations.
The thesis presents methods that encode different types of prior knowledge in the latent space for deep learning methods. The thesis first presents methods that integrate the multi-view geometry into the deep learning frameworks. Then the thesis presents methods that introduce Gaussian Process to capture the correlations of input data in the latent space, which lead to more consistent predictions across frames for applications like depth estimation, stereo matching and novel view synthesis. In addition, the thesis explores two methods that use known factors of variation to learn disentangled representations to increase the controllability for applications like novel view synthesis and interactive image retrieval.
Contact details of the doctoral student: [email protected], +447879511858