Visual Object Class Recognition using Local Descriptions

Huge collections of digital images exist on the web, in large company databases or on our private hard disks. When the images have not been annotated by keywords, finding pictures containing certain objects is a difficult task. This work deals with the recognition of visual object class members in digital photographs, solely using the image data. The fundamental principle of all methods employed and developed in this thesis is the use of local, visual features. One focus is to find suitable prototype parts for modeling the object classes. Several issues like the suitability of interest point detectors, the fast generation of visual codebooks and the discrepancy between visual and semantic similarity of structures are discussed and advanced techniques are developed based on the findings. The second focus of this work is the modeling of the visual object classes using the local parts. In particular, methods that incorporate the geometric relationship of the local features have proven to be beneficial.