Enhancements for local feature based image classification

Using local features with nearest neighbor search and direct voting obtains excellent results for various image classification tasks. In this work we decompose the method into its basic steps which are investigated in detail. Different feature extraction techniques, distance measures, and probability models are proposed and evaluated. We show that improvements are possible for each of the investigated enhancements. This shows that the important aspect of the framework is the decomposition of the training images into sets of local features for each class.