Face tracking and head pose estimation using convolutional neural networks

Asteriadis, S; Karpouzis, K; Kollias, S

dc.contributor.author	Asteriadis, S	en
dc.contributor.author	Karpouzis, K	en
dc.contributor.author	Kollias, S	en
dc.date.accessioned	2014-03-01T02:46:47Z
dc.date.available	2014-03-01T02:46:47Z
dc.date.issued	2010	en
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/32855
dc.subject	Face Tracking	en
dc.subject	Head Pose Estimation	en
dc.subject	Neural Network	en
dc.subject.other	Bounding box	en
dc.subject.other	Convolutional neural network	en
dc.subject.other	Face orientation	en
dc.subject.other	Face Tracking	en
dc.subject.other	Head Pose Estimation	en
dc.subject.other	Proposed architectures	en
dc.subject.other	Rotation angles	en
dc.subject.other	Tracking techniques	en
dc.subject.other	Training data	en
dc.subject.other	Well-aligned	en
dc.subject.other	Animation	en
dc.subject.other	Convolution	en
dc.subject.other	Gesture recognition	en
dc.subject.other	Motion estimation	en
dc.subject.other	Neural networks	en
dc.title	Face tracking and head pose estimation using convolutional neural networks	en
heal.type	conferenceItem	en
heal.identifier.primary	10.1145/1924035.1924046	en
heal.identifier.secondary	http://dx.doi.org/10.1145/1924035.1924046	en
heal.publicationDate	2010	en
heal.abstract	In applications where face orientation is necessary, but in unpretending environments in terms of lighting, equipment, resolution, etc, employing local tracking techniques would usually fail to give accurate results, regarding the issue of head pose estimation. However, in a similar manner, holistic techniques require the face to be well aligned with the training data. This pre-assumes correct and accurate face tracking, which is also a challenging issue. Here, we propose a face tracker, adjusted to each person's face chrominance values, and learnt online. Based on the face bounding box, Convolutional Neural Networks (CNNs) are employed, in order to calculate face orientation. CNNs are ideal for cases where a lot of distortions exist in the data, and the proposed architecture only utilizes subsets of classifiers, excluding those corresponding to rotation angles far from the current.	en
heal.journalName	ACM International Conference Proceeding Series	en
dc.identifier.doi	10.1145/1924035.1924046	en
dc.identifier.spage	19	en