Publication
Using Mutual Independence of Slow Features for Improved Information Extraction and Better Hand-Pose Classification
Aditya Tewari; Bertram Taetz; Frédéric Grandidier; Didier Stricker
In: Václav Skala; et. al. (Hrsg.). Journal of WSCG (WSCG), Vol. 23, No. 1, Pages 35-43, Václav Skala - UNION Agency, Plsen, 7/2015.
Abstract
We propose a Slow Feature Analysis (SFA) based classification of hand-poses and demonstrate that the property of mutual independence of the slow feature functions improves the classification performance. SFA extracts func- tions that describe trends in a time series data and is capable of isolating noise from information while conserving high-frequency components of the data which are consistently present over time or in the set of data points. SFA is a useful knowledge extraction method that can be modified to identify functions which are well suited for distin- guishing classes. We show that by using the orthogonality property of SFA our information about classes can be increased. This is demonstrated by classification results on the well known MNIST dataset for hand written digit detection.
Furthermore, we use a hand-pose dataset with five possible classes to show the performance of SFA. It consistently achieves a detection rate of over 96% for each class. We compare the classification results on shape descrip- tive physical features, on the Principal Component Analysis (PCA) and the non-linear dimensionality reduction (NLDR) for manifold learning. We show that a simple variance based decision algorithm for SFA gives higher recognition rates than K-Nearest Neighbour (KNN), on physical features, PCA and non-linear low dimensional representation. Finally, we examine Convolutional Neural Networks (CNN) in relation with SFA.