Spatial Relations of Features and Descriptors for Appearance Based Object Recognition

Inffeldgasse 16a/2nd floor, A-8010 Graz, Austria

Appearance based object recognition using local features is a very wide area of research. Up to now, a lot of promising approaches have been proposed, still some key issues remain open. There is a requirement for substantial improvements with respect to algorithm runtime, necessary storage space or recognition performance on more comprehensive databases. In this thesis, we investigate simple spatial relations on different levels of the recognition pipeline and propose four novel algorithms to improve the recognition performance of specific object recognition tasks. First we develop a novel distinguished region detector which is complementary to existing approaches. The basic idea is to find distinguished regions by spatially related interest points and we introduce the concept of ‘maximal stableness across scale’. Therefore, the detected regions are called ‘Maximally Stable Corner Clusters (MSCC)’. Second, we design a new descriptor based on the two dimensional joint occurrence histogram of local orientations, which is ideally suited for the regions detected by the MSCCs. In a set of experiments, we demonstrate competitive performance of the developed detector and descriptor approach with respect to other state of the art approaches. Our detector consistently detects regions different from those found by other detectors. Subsequently we present a completely new object representation based only on the binary joint occurrences of quantized descriptors. This representation, despite being high dimensional, is very sparse and efficient. Moreover we can verify, that a high recognition performance is obtained even if we work with very weak, non discriminative descriptors. Finally we propose an efficient method to substantially increase the recognition performance of a standard vocabulary tree based recognition system with our novel binary joint occurrences representation. Extensive experimental evaluations on rather challenging recognition tasks show the robustness of the proposed algorithms to partial occlusions and substantial background clutter. Investigation of the scaling behavior and some profiling experiments finally demonstrate the efficiency and appropriateness of the developed methods to solve specific object recognition tasks.

For a PDF-version of the thesis please contact the author at JOANNEUM RESEARCH.