Cue combination for object recognition

In previous work we have developed a set of regional image descriptors that have been shown to lead to better recognition performance compared to previous histogram descriptors within the same class. In the experimental evaluations, these have, however, been applied in isolation, using only a single type of image descriptor for each recognition task. Moreover, the image descriptors have been computed globally for each image.

An interesting question concerns whether the recognition performance could be improved by combining several such image descriptors in a multi-cue fashion. Possible cue combination techniques include voting and tree-based classifiers. Another issue concerns the computation of multiple regional image descriptors in different subwindows of the image, so as to enable recognition of objects or structures that only occupy a part of the image.

The task of this Master´s thesis project is to investigate the possibilities of these extensions, including the development of cue combination methods and evaluating them on benchmark problems. A natural extension is to investigate the possibility of developing fast and scalable search methods to enable efficient recognition in relation to larger datasets of known objects or images.