## Face Recognition From Long-Term Observations (2002)

### Cached

### Download Links

Venue: | In Proc. IEEE European Conference on Computer Vision |

Citations: | 96 - 2 self |

### BibTeX

@INPROCEEDINGS{Shakhnarovich02facerecognition,

author = {Gregory Shakhnarovich and John W. Fisher and Trevor Darrell},

title = {Face Recognition From Long-Term Observations},

booktitle = {In Proc. IEEE European Conference on Computer Vision},

year = {2002},

pages = {851--868}

}

### Years of Citing Articles

### OpenURL

### Abstract

We address the problem of face recognition from a large set of images obtained over time - a task arising in many surveillance and authentication applications. A set or a sequence of images provides information about the variability in the appearance of the face which can be used for more robust recognition. We discuss di#erent approaches to the use of this information, and show that when cast as a statistical hypothesis testing problem, the classification task leads naturally to an information-theoretic algorithm that classifies sets of images using the relative entropy (Kullback-Leibler divergence) between the estimated density of the input set and that of stored collections of images for each class. We demonstrate the performance of the proposed algorithm on two medium-sized data sets of approximately frontal face images, and describe an application of the method as part of a view-independent recognition system.

### Citations

9150 |
Elements of information theory
- Cover, Thomas
- 1991
(Show Context)
Citation Context ...e i.i.d. assumption it can be shown that the log-likelihood of samples drawn from p l (x) under model distribution p k (x) has the following expected (with respect to p l (x)) and asymptotic behavior =-=[4]-=-: E p l # 1 N N # i=1 log # p k # x i l ## # = - (H(p l (x)) +D (p l (x) ||p k (x))) (7) = lim N## # 1 N N # i=1 log # p k # x i l ## # (8) where D (p l ||p k ) is the well-known asymmetric Kullback-L... |

1636 | Eigenfaces vs. fisherfaces: recognition using class specific linear projection
- Belhumeur, Hespanha, et al.
- 1997
(Show Context)
Citation Context ...n extensively 3 explored [3]. The state of the art today is defined by a family of recognition schemes related to eigendecomposition of the data directly [11] and through Linear Discriminate Analysis =-=[1]-=-, or to local feature analysis of images [13, 17], all of which achieve very high performance. However, face recognition from an image sequence, or more generally from a set of images, has been a subj... |

1286 | Information theory and statistics
- Kullback
- 1968
(Show Context)
Citation Context ... and p k (x) the model density inferred from that sample set. Although not commonly used, the K-ary hypothesis test: H 1 : X (n) 1 # p 0 (x) . . . HK : X (n) K # p 0 (x) (4) is, under mild conditions =-=[10]-=-, equivalent to 3. In 3 we quantify and compare the ability of our inferred model densities to explain the samples under test, while in equation 4 we infer the density of our test samples and then qua... |

1093 | On combining classifiers
- Kittler, Hatef, et al.
- 1998
(Show Context)
Citation Context ...tion where multiple observations of the same person's face are accumulated over time. This is an instance of the more general problem of fusion of evidence from multiple measurements. Kittler et alin =-=[9]-=- present a statistical interpretation of a number of common methods for cross- modal fusion, such as the product, maximum , and majority rules, which are also appropriate for late integration over a s... |

882 | Robust real-time object detection
- Viola, Jones
- 2004
(Show Context)
Citation Context ...sation; we refer to this data set as the Conversation data. Each person is represented by one clip captured with a CCD camera, at 30 fps. The faces were detected using a fast face detection algorithm =-=[16]-=-, and resized to 2222 pixel gray level images. False positives were manually removed. In all of the images, the pose is frontal within about 15 degrees, with an even smaller range of tilt and yaw. The... |

600 | Probabilistic visual learning for object representation
- Moghaddam, Pentland
(Show Context)
Citation Context ...d its relationship to the Kullback-Leibler divergence between distributions. In reality the distributions p k are unknown and need to be estimated from data, as well as p 0 . In this paper, we follow =-=[12]-=- and estimate the densities in the space of frontal face images by a multivariate Gaussian. Each subject has its own 6 density, which is estimated based on the training samples of that subject's face.... |

560 | W4: real-time surveillance of people and their activities
- Haritaoglu, Harwood, et al.
- 2000
(Show Context)
Citation Context ...n developed (e.g. Visionics, Bochum/USC). Separately, systems for tracking people in unconstrained environments have become increasingly robust, and are able to track individuals for minutes or hours =-=[8, 5]-=-. These systems typically can provide images of users at low or medium resolution, possibly from multiple viewpoints, over long periods of time. For optimal recognition performance, the information fr... |

478 |
von der Malsburg. Face Recognition by Elastic
- Wiskott, Fellous, et al.
- 1997
(Show Context)
Citation Context ...he art today is defined by a family of recognition schemes related to eigendecomposition of the data directly [11] and through Linear Discriminate Analysis [1], or to local feature analysis of images =-=[13, 17]-=-, all of which achieve very high performance. However, face recognition from an image sequence, or more generally from a set of images, has been a subject of relatively few published studies. The comm... |

241 | Local Feature Analysis: A general statistical theory for object representation
- Penev, Atick
- 1996
(Show Context)
Citation Context ...he art today is defined by a family of recognition schemes related to eigendecomposition of the data directly [11] and through Linear Discriminate Analysis [1], or to local feature analysis of images =-=[13, 17]-=-, all of which achieve very high performance. However, face recognition from an image sequence, or more generally from a set of images, has been a subject of relatively few published studies. The comm... |

232 |
Statistical Signal Processing, Detection Estimation and Time Series Analysis
- Scharf
- 1990
(Show Context)
Citation Context ...in our model samples. Assuming that the x i are i.i.d. (independent and indentically distributed) and the classes have equal prior probability, it is well known, by the Neyman-Pearson lemma (e.g. see =-=[14]-=-), that an optimal statistic for choosing one hypothesis over the other is the log-likelihood function; that is, H k = argmax k # i log # p k # x i 0 ## (5) for the hypothesis test of (3) and H k = ar... |

128 |
Bayesian face recognition
- Moghaddam, Jebara, et al.
- 2000
(Show Context)
Citation Context ...ed and global, template-based methods have been extensively 3 explored [3]. The state of the art today is defined by a family of recognition schemes related to eigendecomposition of the data directly =-=[11]-=- and through Linear Discriminate Analysis [1], or to local feature analysis of images [13, 17], all of which achieve very high performance. However, face recognition from an image sequence, or more ge... |

101 |
ace Recognition Using Temporal Image Sequence
- Yamaguchi, Fukui, et al.
- 1998
(Show Context)
Citation Context ...s classification of X (n) based on the mean imagesx. More sophisticated approaches capture higher-order moments of the distribution. In particular, the Mutual Subspace Method (MSM) of Yamaguchi et al =-=[18]-=- is noteworthy. In MSM, a test set of images is represented by the linear subspace spanned by the principal components of the data. This subspace is then compared to each of the subspaces constructed ... |

69 |
Face recognition: features vs. templates
- Brunelli
- 1993
(Show Context)
Citation Context ... in Section 7. 2 Previous work The area of recognition from a single face image is very well established. Both local, feature-based and global, template-based methods have been extensively 3 explored =-=[3]-=-. The state of the art today is defined by a family of recognition schemes related to eigendecomposition of the data directly [11] and through Linear Discriminate Analysis [1], or to local feature ana... |

54 | Intergrated Face and Gait Recognition From Multiple Views
- Shakhnarovich, Lee, et al.
- 2001
(Show Context)
Citation Context ...st was a collection of people observed at a distance in frontal view in an indoor o#ce environment. The second was computed using a recently proposed scheme for integration across multiple viewpoints =-=[15]-=-, in which images from several cameras with arbitrary views of the user's face were combined to generate a virtual frontal view. With both data sets, our distribution matching technique o#ered equal o... |

47 | Plan-view trajectory estimation with dense stereo background models
- Darrell, Demirdjian, et al.
- 2001
(Show Context)
Citation Context ...n developed (e.g. Visionics, Bochum/USC). Separately, systems for tracking people in unconstrained environments have become increasingly robust, and are able to track individuals for minutes or hours =-=[8, 5]-=-. These systems typically can provide images of users at low or medium resolution, possibly from multiple viewpoints, over long periods of time. For optimal recognition performance, the information fr... |

25 |
Cootes, “Improving Identification Performance by Integrating Evidence from Sequences
- Edwards, Taylor, et al.
(Show Context)
Citation Context ...e variability in the observed data. Recently proposed detection and tracking algorithms use the dynamics of the consecutive images of a face, and integrate recognition into the tracking framework. In =-=[6], the vari-=-ation of the individual faces is modeled in the framework of the Active Appearance Model. In [7], the "temporal signature" of a face is defined, and feed-forward neural network is used in or... |

17 | Tracking and recognition of face sequences
- Gong, Psarrou, et al.
- 1994
(Show Context)
Citation Context ...f the consecutive images of a face, and integrate recognition into the tracking framework. In [6], the variation of the individual faces is modeled in the framework of the Active Appearance Model. In =-=[7], the &quo-=-t;temporal signature" of a face is defined, and feed-forward neural network is used in order to classify the sequence. In [2], people are recognized from a sequence of rotating head images, using... |

14 | Face Recognition from Multi-Pose Image Sequence
- Biuk, Loncaric
- 2001
(Show Context)
Citation Context ... faces is modeled in the framework of the Active Appearance Model. In [7], the "temporal signature" of a face is defined, and feed-forward neural network is used in order to classify the seq=-=uence. In [2]-=-, people are recognized from a sequence of rotating head images, using trajectories in a low-dimensional eigenspace. They assume that each image is associated with a known pose -- a situation which is... |

9 |
Dual differential geometry associated with the Kullback-Leibler information on the Gaussian distributions and its 2-parameter deformations
- Yoshizawa, Tanabe
- 1999
(Show Context)
Citation Context ...h is closest to the “true” test distribution and then, of the training distributions, selecting that which is closest to the estimated testing density in the K-L sense. Finally, using a recent result =-=[19]-=-, we estimate Gaussian distribution parameters for both training and test data and compute K-L divergences analytically. 5 Classification based on Kullback-Leibler divergence In this section we define... |

1 |
Dual di#erential geometry associated with the Kullback-Leibler information on the Gaussian distributions and its 2-parameter deformations
- Yoshizawa, Tanabe
- 1999
(Show Context)
Citation Context ...h is closest to the "true" test distribution and then, of the training distributions, selecting that which is closest to the estimated testing density in the K-L sense. Finally, using a rece=-=nt result [19]-=-, we estimate Gaussian distribution parameters for both training and test data and compute K-L divergences analytically. 5 Classification based on Kullback-Leibler divergence In this section we define... |