Disputation Katharina Schwarz
am Montag, 23. Juli 2018, um 13:00 Uhr in Raum A 301, Sand 1
Text–to–Video: Image Semantics and NLP
Berichterstatter 1: Prof. Dr. Hendrik P. A. Lensch
Berichterstatter 2: Prof. Dr. Andreas Schilling
Kurze Zusammenfassung des Vortrags
When aiming at automatically translating an arbitrary text into a visual story, the main challenge consists in finding a semantically close visual representation whereby the displayed meaning should remain the same as in the given text. Besides, the appearance of an image itself largely influences how its meaningful information is transported towards an observer. This thesis now demonstrates that investigating in both, image semantics as well as the semantic relatedness between visual and textual sources enables us to tackle the challenging semantic gap and to find a semantically close translation from natural language to a corresponding visual representation.
Within the last years, social networking became of high interest leading to an enormous and still increasing amount of online available data. Photo sharing sites like Flickr allow users to associate textual information with their uploaded imagery. Thus, this thesis exploits this huge knowledge source of user generated data providing initial links between images and words, and other meaningful data.
In order to approach visual semantics, this work presents various methods to analyze the visual structure as well as the appearance of images in terms of meaningful similarities, aesthetic appeal, and emotional affect towards an observer. In detail, our GPU-based approach efficiently finds visual similarities between images in large datasets across visual domains and identifies various meanings for ambiguous words exploring similarity in online search results. Further, we investigate in the highly subjective aesthetic appeal of images and make use of deep learning to directly learn aesthetic rankings from a broad diversity of user reactions in social online behavior. To gain even deeper insights into the influence of visual appearance towards an observer, we explore how simple image processing is capable of actually changing the emotional perception and derive a simple but effective image filter.
To identify meaningful connections between written text and visual representations, we employ methods from Natural Language Processing (NLP). Extensive textual processing allows us to create semantically relevant illustrations for simple text elements as well as complete storylines. More precisely, we present an approach that resolves dependencies in textual descriptions to arrange 3D models correctly. Further, we develop a method that finds semantically relevant illustrations to texts of different types based on a novel hierarchical querying algorithm. Finally, we present an optimization based framework that is capable of not only generating semantically relevant but also visually coherent picture stories in different styles.