Semantic interest point detection

In this thesis, the problem of constructing a generic model-driven visual dictionary that is adequate for a variety of image understanding applications is investigated. The proposed dictionary based scheme that we call Symbolic Patch Dictionary (SymPaD), follows the steps of Bag of Visual Words (BoVW) paradigm in which, pixels are visited on a dense grid, local image characteristics are extracted in terms of shape similarity scores to the dictionary atoms, the scores are pooled, and finally an image signature is obtained. We differ from BoVW schemes in the literature in the generation of our shape dictionary. These shape patterns are generated by mathematical formulae encoding qualititative image characteristics. Compared with the existing model-driven schemes our method is able to represent images of a variety benchmark datasets of image understanding applications with better discrimination.