Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion

In supervised learning of head pose classification, uniformly distributed and labeled ground-truth of people in large quantities is required. Unfortunately, labeling is both tedious and inconsistent. As a result, the classifier could not generalize well the unseen data. To address this problem, in t...

Full description

Saved in:
Bibliographic Details
Main Authors: Wang, Jian-Gang., Yau, Wei-Yun., Sung, Eric.
Other Authors: School of Electrical and Electronic Engineering
Format: Conference or Workshop Item
Language:English
Published: 2013
Subjects:
Online Access:https://hdl.handle.net/10356/102835
http://hdl.handle.net/10220/16874
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
id sg-ntu-dr.10356-102835
record_format dspace
spelling sg-ntu-dr.10356-1028352020-03-07T13:24:51Z Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion Wang, Jian-Gang. Yau, Wei-Yun. Sung, Eric. School of Electrical and Electronic Engineering IAPR International Conference on Biometrics (5th : 2012 : New Delhi, India) DRNTU::Engineering::Electrical and electronic engineering In supervised learning of head pose classification, uniformly distributed and labeled ground-truth of people in large quantities is required. Unfortunately, labeling is both tedious and inconsistent. As a result, the classifier could not generalize well the unseen data. To address this problem, in this paper we propose a novel bootstrapping (semi-supervised) which can train a classifier using both a small number of labeled data and an abundance of unlabeled data. The difficulty of using unlabeled data is that the performance could be worse than using just the labeled data if the unlabeled data has a different feature space distribution than that of the labeled data. This could be the reason why there is little work done to estimate the head pose with unlabeled data. In our proposed method, automatic data mining is applied to select unlabeled data that has higher likelihood to be helpful to improve the performance of a classifier trained solely on the labeled data. Kernel linear discriminant analysis and a SIFT-based image registration are combined to predict the head pose from face image. Some pose prototypes, learned from the labeled samples, are used to define a novel confidence measurement for selecting the unlabeled data. Experimental results on a large database verified that the proposed bootstrapped approach can achieve significantly better performance than the supervised learning alone. 2013-10-25T02:01:13Z 2019-12-06T21:00:56Z 2013-10-25T02:01:13Z 2019-12-06T21:00:56Z 2012 2012 Conference Paper Wang, J. G., Yau, W. Y., & Sung, E. (2012). Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion. 2012 5th IAPR International Conference on Biometrics (ICB), 32-39. https://hdl.handle.net/10356/102835 http://hdl.handle.net/10220/16874 10.1109/ICB.2012.6199755 en
institution Nanyang Technological University
building NTU Library
country Singapore
collection DR-NTU
language English
topic DRNTU::Engineering::Electrical and electronic engineering
spellingShingle DRNTU::Engineering::Electrical and electronic engineering
Wang, Jian-Gang.
Yau, Wei-Yun.
Sung, Eric.
Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
description In supervised learning of head pose classification, uniformly distributed and labeled ground-truth of people in large quantities is required. Unfortunately, labeling is both tedious and inconsistent. As a result, the classifier could not generalize well the unseen data. To address this problem, in this paper we propose a novel bootstrapping (semi-supervised) which can train a classifier using both a small number of labeled data and an abundance of unlabeled data. The difficulty of using unlabeled data is that the performance could be worse than using just the labeled data if the unlabeled data has a different feature space distribution than that of the labeled data. This could be the reason why there is little work done to estimate the head pose with unlabeled data. In our proposed method, automatic data mining is applied to select unlabeled data that has higher likelihood to be helpful to improve the performance of a classifier trained solely on the labeled data. Kernel linear discriminant analysis and a SIFT-based image registration are combined to predict the head pose from face image. Some pose prototypes, learned from the labeled samples, are used to define a novel confidence measurement for selecting the unlabeled data. Experimental results on a large database verified that the proposed bootstrapped approach can achieve significantly better performance than the supervised learning alone.
author2 School of Electrical and Electronic Engineering
author_facet School of Electrical and Electronic Engineering
Wang, Jian-Gang.
Yau, Wei-Yun.
Sung, Eric.
format Conference or Workshop Item
author Wang, Jian-Gang.
Yau, Wei-Yun.
Sung, Eric.
author_sort Wang, Jian-Gang.
title Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
title_short Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
title_full Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
title_fullStr Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
title_full_unstemmed Head pose estimation by bootstrapping generalized discriminant analysis with SIFT flow alignment criterion
title_sort head pose estimation by bootstrapping generalized discriminant analysis with sift flow alignment criterion
publishDate 2013
url https://hdl.handle.net/10356/102835
http://hdl.handle.net/10220/16874
_version_ 1681041374492753920