Reliable gesture recognition with transductivie confidence machines

Abstract

The transductive confidence machines (TCMs) framework allows to extend classifiers such that their performance can be set by the user prior to classification. In this chapter we briefly survey different approaches of using the TCM framework. Most applications of TCM are constrained to relatively few data samples with a limited number of classes, due to the computational complexity of the TCM approach. A novel technique is presented for reducing the computational costs and memory consumption required for updating the non-conformity scores in the offline learning setting of TCMs. The improved TCM, using a k-nearest neighbor classifier, is evaluated by applying it to the NicIcon collection of iconic gestures, acquired in the critical domain of crisis management. For such domains, reliable classification is very important. The results show that TCMs outperform previous methods on this dataset, on both relatively easy data and on difficult test samples.