An efficient probabilistic online classification approach for object recognition with random forests

Abstract

Online learning on big data sets is still an open problem in the classification of images. Many problems in the real world don't have all data available in the beginning of the training. Therefore it is necessary that the approach is able to integrate new incoming datapoints. Random Forest have been proven to be good in online learning. However the existing approaches do only generate very few trees, which only have a height of five. To overcome this shortcoming this thesis presents several methods to improve the generation of Decision trees, which leads to an algorithm, which can train thousands of tree with a sufficient height. Furthermore the Random Forest were then used in combination with an online sparse Gaussian Process to classify the outliners. These falsely classified points weren't classified correctly by the Random Forest in the first place. This whole approach was then optimized and tested on different datasets. The far most important result was that the presented online approach always yields better results than the offline approach, which is a remarkable result for an online learning approach. Furthermore we outperformed the result from Saffari et al. on the USPS dataset.