<XML><RECORDS><RECORD><REFERENCE_TYPE>3</REFERENCE_TYPE><REFNUM>8256</REFNUM><AUTHORS><AUTHOR>Lo,T.W.R.</AUTHOR><AUTHOR>Siebert,J.P.</AUTHOR><AUTHOR>Ayoub,A.F.</AUTHOR></AUTHORS><YEAR>2006</YEAR><TITLE>Robust Feature Extraction for Range Images Interpretation using Local Topology Statistics</TITLE><PLACE_PUBLISHED>Proceedings of MICCAI 2006 Workshop on Craniofacial Image Analysis for Biology, Clinical Genetics, Diagnostics and Treatment</PLACE_PUBLISHED><PUBLISHER>N/A</PUBLISHER><PAGES>75-82</PAGES><LABEL>Lo:2006:8256</LABEL><ABSTRACT>This paper presents a robust methodology for extracting feature vectors, which exhibit invariance to viewpoint rotations, from 2.5-dimensional (2.5D) range images of human faces. Our approach is based on extracting the topology statistics within the local surface sampled at an anatomic landmark location. These statistics are used to define a feature vector that encapsulates the shape of the sampled surface by means of the relative frequencies of classified surface types and their orientations. The invariance of this feature vector against rotational changes in viewpoint for all three Euler axes has been validated using 50 frontal range images of human faces, each labelled with 28 anatomic landmarks. Our results indicate that the invariance of the feature vectors to in-plane and out-of-plane rotations of $<=$ 80 degrees, remained on average, within the correlation range 0.7-0.9, as measured by the vector dot product. Under these conditions a subset of 15 landmarks remained usefully discriminable.</ABSTRACT></RECORD></RECORDS></XML>