Spatial Pyramids and Two-layer Stacking SVM classifiers for Image Categorization: A Comparative Study

Recent research in image recognition has shownthat combining multiple descriptors is a very useful way toimprove classification performance. Furthermore, the use ofspatial pyramids that compute descriptors at multiple spatialresolution levels generally increases the discriminative power ofthe descriptors. In this paper we focus on combination methodsthat combine multiple descriptors at multiple spatial resolutionlevels. A possible problem of the naive solution to create onelarge input vector for a machine learning classifier such asa support vector machine, is that the input vector becomesof very large dimensionality, which can increase problems ofoverfitting and hinder generalization performance. Thereforewe propose the use of stacking support vector machines whereat the first layer each support vector machine receives the inputconstructed by each single descriptor and is trained to computethe right output class. A second layer support vector machineis then used to combine the class probabilities of all trainedfirst layer support vector models to learn the right outputclass given these reduced input vectors. We have performedexperiments on 20 classes from the Caltech object database with10 different single descriptors at 3 different resolutions. Theresults show that our 2-layer stacking approach outperformsthe naive approach that combines all descriptors directly in avery large single input vector