@inproceedings {pub2554,
	title = {Beyond Histograms: Why Learned Structure-Preserving Descriptors Outperform Statistical Features},
	author = {Thomas Guthier AND Volker Willert AND Julian Eggert},
	year = {2014},
	month = {April},
	abstract = {Statistical image descriptors based on histograms~\cite{lowe2004distinctive,dalal2006} are widely used in image processing, because they are fast and simple methods with high classification performance. However, they discard the local topological structure and thus lose descriminative information contained in the image. We show that features based on the simple cell/complex cell response of learned structure-preserving patterns outperform the statistical HOG descriptors for two visual human action recognition datasets~\cite{blank2005,ucfSports}. The patterns are learned with a shift-invariant, sparse, non-negative unsupervised learning algorithm~\cite{eggert2003,eggert2004,guthier2012}, that provides a distinct decomposition of the input into its parts. Alongside with an increasing number of patterns, the discriminative power of the features and thus the classification performance is increasing.},
	publisher = {ESANN},
	booktitle = {European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN) 2014},
	city = {Bruges, Belgium}
}