Training Support Vector Machines on Large Sets of Image Data

Abstract

Object detection problems in computer vision often present a computationally difficult task in machine learning, where very large amounts of high-dimensional image data have to be processed by complex training algorithms. We consider training support vector machine (SVM) classifiers on big sets of image data and investigate approximate decomposition techniques that can use any limited conventional SVM training tool to cope with large training sets. We reason about expected comparative performance of different approximate training schemes and subsequently suggest two refined training algorithms, one aimed at maximizing the accuracy of the resulting classifier, the other allowing very fast and rough preview of the classifiers that can be expected from given training data. We show how the best approximation method trained on an augmented training set of one million perturbed data samples outperforms an SVM trained on the original set.