Use this URL to cite or link to this record in EThOS:
Title: Synthesizing and editing photo-realistic visual objects
Author: Turmukhambetov, D.
ISNI:       0000 0004 7231 2214
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2016
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
In this thesis we investigate novel methods of synthesizing new images of a deformable visual object using a collection of images of the object. We investigate both parametric and non-parametric methods as well as a combination of the two methods for the problem of image synthesis. Our main focus are complex visual objects, specifically deformable objects and objects with varying numbers of visible parts. We first introduce sketch-driven image synthesis system, which allows the user to draw ellipses and outlines in order to sketch a rough shape of animals as a constraint to the synthesized image. This system interactively provides feedback in the form of ellipse and contour suggestions to the partial sketch of the user. The user's sketch guides the non-parametric synthesis algorithm that blends patches from two exemplar images in a coarse-to-fine fashion to create a final image. We evaluate the method and synthesized images through two user studies. Instead of non-parametric blending of patches, a parametric model of the appearance is more desirable as its appearance representation is shared between all images of the dataset. Hence, we propose Context-Conditioned Component Analysis, a probabilistic generative parametric model, which described images with a linear combination of basis functions. The basis functions are evaluated for each pixel using a context vector computed from the local shape information. We evaluate C-CCA qualitatively and quantitatively on inpainting, appearance transfer and reconstruction tasks. Drawing samples of C-CCA generates novel, globally-coherent images, which, unfortunately, lack high-frequency details due to dimensionality reduction and misalignment. We develop a non-parametric model that enhances the samples of C-CCA with locally-coherent, high-frequency details. The non-parametric model efficiently finds patches from the dataset that match the C-CCA sample and blends the patches together. We analyze the results of the combined method on the datasets of horse and elephant images.
Supervisor: Kautz, J. ; Prince, S. J. D. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available