Inference on highly-connected discrete graphical models with applications to visual object recognition [Elektronische Ressource] / vorgelegt von Jörg Hendrik Kappes

Inaugural-DissertationzurErlangung der DoktorwurdederNaturwissenschaftlich{Mathematischen GesamtfakultatderRuprecht{Karls{UniversitatHeidelbergvorgelegt vonDipl.{Inf. J org Hendrik Kappesaus HeidelbergTag der mundlic hen Prufung: 18.04.2011Inference on Highly-Connected DiscreteGraphical Models with Applications to VisualObject RecognitionGutachter: Prof. Dr. Christoph Schn orrProf. Dr. Gerhard ReineltAbstractObject detection is one of the key components of modern computer vision systems. Whilethe detection of a speci c rigid object under changing viewpoints was considered hardjust a few years ago, current research strives to detect and recognize classes of non-rigid,articulated objects. Hampered by the omnipresent problems due to clutter and occlu-sion, the focus has shifted from holistic approaches for object detection to representationsof individual object parts linked by structural information, along with richer contextualdescriptions of object con gurations.Along this line of research, we present a practicable and expandable probabilistic frame-work for parts-based object class representation, using probabilistic graphical models,enabling the detection of rigid and articulated object classes in arbitrary views. We in-vestigate computational learning of this representation from labelled training images andinfer globally optimal solutions to the contextual maximum a posteriori (MAP) detectionproblem for object recognition.