ausblenden:
Schlagwörter:
-
Zusammenfassung:
In the context of image and video editing, this thesis proposes methods for
modifying the semantic content of a recorded scene. Two different editing
problems are approached: First, the removal of ghosting artifacts from high
dynamic range (HDR) images recovered from exposure sequences, and second, the
removal of objects from video sequences recorded with and without camera
motion. These editings need to be performed in a way that the result looks
plausible to humans, but without having to recover detailed models about the
content of the scene, e.g. its geometry, reflectance, or illumination. The
proposed editing methods add new key ingredients, such as camera noise models
and global optimization frameworks, that help achieving results that surpass
the capabilities of state-of-the-art methods. Using these ingredients, each
proposed method defines local visual properties that approximate well the
specific editing requirements of each task. These properties are then encoded
into a energy function that, when globally minimized, produces the required
editing results. The optimization of such energy functions corresponds to
Bayesian inference problems that are solved efficiently using graph cuts. The
proposed methods are demonstrated to outperform other state-of-the-art methods.
Furthermore, they are demonstrated to work well on complex real-world scenarios
that have not been previously addressed in the literature, i.e., highly
cluttered scenes for HDR deghosting, and highly dynamic scenes and
unconstrained camera motion for object removal from videos.