Item

Abstract

Correspondence problems like optic flow belong to the fundamental problems in
computer vision. Here, one aims at finding correspondences between the pixels
in two (or more) images. The correspondences are described by a displacement
vector field that is often found by minimising an energy (cost) function. In
this thesis, we present several contributions to the energy-based solution of
correspondence problems: (i) We start by developing a robust data term with a
high degree of invariance under illumination changes. Then, we design an
anisotropic smoothness term that works complementary to the data term, thereby
avoiding undesirable interference. Additionally, we propose a simple method for
determining the optimal balance between the two terms. (ii) When discretising
image derivatives that occur in our continuous models, we show that adopting
one-sided upwind discretisations from the field of hyperbolic differential
equations can be beneficial. To ensure a fast solution of the nonlinear system
of equations that arises when minimising the energy, we use the recent fast
explicit diffusion (FED) solver in an explicit gradient descent scheme. (iii)
Finally, we present a novel application of modern optic flow methods where we
align exposure series used in high dynamic range
(HDR) imaging. Furthermore, we show how the alignment information can be used in
a joint super-resolution and HDR method