Kernel Based Learning of Distances for Shape Detection

In this work we investigate kernel methods for object detection. Kernel principal component analysis (KPCA) is the algorithm used to generate nonlinear models of object variation. We focus on edge-based features, i.e. shapes, to represent objects. Object variations are learned a priori in feature space. Based on the Mahalanobis distance we define a distance measure for an object. In contrast to other works done on shape distances with KPCA, our methods are alignment free. Distance transformed shapes are employed to achieve invariance against small variations in object appearance. The “bag of tuples” approach is used to derive a totaly alignment free, parametric shape representation. Finally we test our approaches. The approaches are evaluated using their receiver operator characteristic (ROC).