#### DMCA

## Robust Parameterized Component Analysis: Theory and Applications to 2D Facial Modeling (2002)

### Cached

### Download Links

Venue: | Computer Vision and Image Understanding, 91:53 – 71 |

Citations: | 53 - 12 self |

### Citations

3778 | Eigenfaces for recognition.
- Turk, Pentland
- 1991
(Show Context)
Citation Context ...lving a linear system of equations. 2.2 Adding motion into the subspace formulation Since the preliminary work of Sirovich and Kirby [48] and the successful eigenface application of Turk and Pentland =-=[49]-=-, PCA has been widely applied to the construction of a face subspace. Since then, there has been a lot of work and interest in trying to construct more accurate models of the high dimensional manifold... |

2144 |
An Introduction to Genetic Algorithms.
- Mitchell
- 1998
(Show Context)
Citation Context ...e can move more than 20 pixels from frame to frame). In order to cope with such real conditions, we explore the use of stochastic methods such as Simulated Annealing (SA) [2], Genetic Algorithms (GA) =-=[27,29]-=- or Condensation (particle filtering) [6,17] for motion estimation. Although the techniques are very similar computationally speaking, here we make use of GA [29] within a coarse-to-fine strategy. Giv... |

2123 | Active appearance models,”
- Cootes, Edwards, et al.
- 2001
(Show Context)
Citation Context ...deo-conferencing, realistic avatar animation and eye tracking. 1 Introduction Many computer vision researchers have used Principal Component Analysis (PCA) to parameterize appearance, shape or motion =-=[3,10,23,30]-=-. However, one major drawback of this traditional technique is that it needs normalized samples in the training data. In the case of computer vision applications, the result is that the samples have t... |

934 | Robust Statistics: The Approach Based on Influence Functions. - Hampel, Ronchetti, et al. - 1986 |

888 | [Visual Reconstruction],
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...der to reduce the influence of outlying data. σp is a parameter that controls p the convexity of the robust function and is used for the deterministic annealing in a Graduated Non-Convexity algorithm=-= [4,7] -=-(we do not minimize Erereg over σp). Benefits of the robust formulation in the subspace related problems are explained elsewhere [14]. Observe that the previous Eq. (4) is similar to Eigentracking [4... |

828 | Detecting faces in images: a survey,
- Yang, Kriegman, et al.
- 2002
(Show Context)
Citation Context ...s motion and the lack of labeled points for solving the correspondence problem between frames. Although many computer vision researchers have used principal component analysis (PCA) to model the face =-=[11,17{19,26,40,41,52]-=- the major drawback of this traditional technique is that it requires normalized (aligned) samples in the training data. While, in the recognition process, alignment of the data with respect to the fa... |

768 | View-based and modular eigenspaces for face recognition,
- Pentland, Moghaddam, et al.
- 1994
(Show Context)
Citation Context ...ting the data while aligning the training data with that subspace. To illustrate the method we develop it in the context of face modeling. In particular, we adopt the idea of modular eigenspaces (ME) =-=[31,40,44]-=- and apply our parameterized component analysis technique to the problem of developing person-specic facial appearance models (PSFAM). Consider the problem of learning a linear subspace representing ... |

696 | Probabilistic visual learning for object representation.
- Moghaddam, Pentland
- 1997
(Show Context)
Citation Context ...deo-conferencing, realistic avatar animation and eye tracking. 1 Introduction Many computer vision researchers have used Principal Component Analysis (PCA) to parameterize appearance, shape or motion =-=[3,10,23,30]-=-. However, one major drawback of this traditional technique is that it needs normalized samples in the training data. In the case of computer vision applications, the result is that the samples have t... |

687 |
Active Contours.
- Blake, Isard
- 1998
(Show Context)
Citation Context ...frame). In order to cope with such real conditions, we explore the use of stochastic methods such as Simulated Annealing (SA) [2], Genetic Algorithms (GA) [27,29] or Condensation (particle filtering) =-=[6,17]-=- for motion estimation. Although the techniques are very similar computationally speaking, here we make use of GA [29] within a coarse-to-fine strategy. Given the first image of the sequence, we manua... |

661 | Hierarchical model-based motion estimation.
- Bergen, Anandan, et al.
- 1992
(Show Context)
Citation Context ...rix which contains the p th row of the matrixW of the region l. Minimizing Equation (4) is a non-linear optimization problem w.r.t. the motion parameters. Following previous work on motion estimation =-=[2,4,27]-=-, we linearize the variation of the function, using a 1 st order Taylor series approximation. Without loss of generality, rather than linearizing the transformation which warps the eigenspace towards ... |

594 |
Low-dimensional procedure for the characterization of human faces.
- Sirovich, Kirby
- 1987
(Show Context)
Citation Context ... the bases B with Csxed. Typically, both updates are computed by solving a linear system of equations. 2.2 Adding motion into the subspace formulation Since the preliminary work of Sirovich and Kirby =-=[48]-=- and the successful eigenface application of Turk and Pentland [49], PCA has been widely applied to the construction of a face subspace. Since then, there has been a lot of work and interest in trying... |

556 | efficient region tracking with parametric models of geometry and illumination,
- Hager, Belhumeur
- 1998
(Show Context)
Citation Context ...tic. In this paper we focus on the application of face modeling. Most of the previous work on face tracking and modeling is focused on generic trackers, which are independent of the person’s identit=-=y [5,9,10,23,24]. -=-In particular, appearance based face trackers [10,26,sRobust Parameterized Component Analysis 655 34] make use of PCA in order to construct a linear model of the face’s subspace (variation across pe... |

356 | Statistical Models of Appearance for Computer Vision.
- COOTES, TAYLOR
- 2001
(Show Context)
Citation Context ...and, without a good starting point, gradient descent methods are likely to get trapped in local minima. When computing the motion parameters, as in the case of optical flow, a coarse-to-fine strategy =-=[4,11]-=- can help to avoid local minima. Although a coarse-to-fine strategy is helpful, this technique is insufficient in our case, since in real image sequences the size of the face can be small in compariso... |

296 | Shadow puppetry.
- Brand
- 1999
(Show Context)
Citation Context .... In general it is hard to model and animate faces, even when they are cartoon characters. Usually complex models encoding the physical underlying musculature of the face are used (e.g. Candide model =-=[8]-=-). We use a PSFAM to parameterize the expression using modular PCA, and parameterized component analysis to learn the PSFAM. Fig. 8 shows frames of a virtual female face animated by the appearance of ... |

281 | Local feature analysis: a general statistical theory for object representation,
- Penev, Atick
- 1996
(Show Context)
Citation Context ...to occlusions [37], etc.). However, it is worth pointing out that representations other than ME have been explored successfully for face recognition and tracking; for instance, Local Feature Analysis =-=[43,42]-=- or Gabor jets with elastic graph matching [51]. Although these techniques have shown good performance in recognition and tracking domains, they do not address the issue of learning a model invariant ... |

253 |
Robust regression using iteratively reweighted least-squares,
- Holland, Welsch
- 1977
(Show Context)
Citation Context ... conveniently) we can reformulate the minimization problem as one of iteratively reweighted least-squares (IRLS), which provides an approximate, iterative, solution to the robust M-estimation problem =-=[30,35]-=-. For a given , a matrixW 2 < dT , which contains the positive weights for each pixel and each image, is calculated for each iteration as a function of the previous residuals e pi = d pt ( l p P ... |

241 | Support vector tracking
- Avidan
- 2004
(Show Context)
Citation Context ...learning by simply replacing the closed form solution with a gradient descent algorithm or any adaptive method. Based on the recent extension of EigenTracking [4] to deal with Support Vector Machines =-=[1]-=- it would also be interesting to consider extending our method to other statistical learning techniques like SVM or independent component analysis.s668 F. De la Torre and M.J. Black Modeling the face ... |

232 | Cootes. Automatic interpretation and coding of face images using flexible models
- Lanitis, Taylor, et al.
- 1997
(Show Context)
Citation Context ...ating a subspace for the appearance variation. 2.3 Person specic models While most work on face tracking focuses on generic trackers which are independent of the identity of the person being tracked =-=[5,8,10,11,26,27,34]-=-, here we focus on Person Specic Facial Appearance Models (PSFAM) [17,26,22,50] for tracking a single individual and use PCA to model the variations due to changes in expression. Although PSFAM are o... |

209 |
Statistical methods for tomographic image reconstruction
- Geman, McClure
- 1987
(Show Context)
Citation Context ...p j=1 blpjcl �� jt f1(xp, al t) � � ,σp (4) where bl pj is the pth pixel of the jth basis of Bl for the layer l. Observe that the pixel residual is filtered by the Geman-McClure robust error =-=function [22] giv-=-en by ρ(x, σp) = x 2 x2 +σ2 , in order to reduce the influence of outlying data. σp is a parameter that controls p the convexity of the robust function and is used for the deterministic annealing ... |

207 | Recognizing imprecisely localized, partially occluded, and expression variant faces from a single sample per class - Martinez |

190 | A New Algorithm for Non-Rigid Point Matching
- Chui, Rangarajan
- 2000
(Show Context)
Citation Context ... the image sequence and in some cases the features or curves are placed by hand. If this is the case, the problem is how to put the features in correspondence using rigid or non-rigid transformations =-=[9,29]-=-. In the other direction, Walker et al. [32] have proposed a method for automatically placing landmarks to dene correspondence between images and hence automatically constructing appearance models. S... |

188 | ªRecognizing Facial Expressions in Image Sequences Using Local Parameterized Models of Image Motion,º Int'l
- Black, Yacoob
- 1997
(Show Context)
Citation Context ...s the basis to the image. Observe that f will be approximately the inverse of f1. Recall that ||d|| 2 W = dT Wd is a weighted norm. Wt ∈ℜ d×d is a diagonal matrix, such that the diagonal elements=-= are (5) (6)s-=-662 F. De la Torre and M.J. Black the t th column of W. W l t ∈ℜ dl×dl is diagonal matrix, where the diagonal is created by the elements of the t column of W which belong to the l th layer. Obser... |

174 | A framework for robust subspace learning, - Torre, Black - 2003 |

144 | Principal Component Neural Networks: Theory and Applications - Diamantaras, Kung - 1996 |

133 | Robust principal component analysis for computer vision
- Torre, Black
- 2001
(Show Context)
Citation Context ...ponents (B) can be achieved by computing the k largest eigenvectors of the covariance matrix DD T [18], here it is useful to exploit work that formulates PCA as the minimization of an energy function =-=[14,18]-=-. Related formulations have been studied in various communities 1 Bold capital letters denote a matrix D, bold lower-case letters a column vector d. dj represents the j-th column of the matrix D and d... |

95 | A multi-view nonlinear active shape model using kernal PCA
- Romdhani, Gong, et al.
- 1999
(Show Context)
Citation Context ...ring the last few years there has been a growing trend to apply new machine learning or multivariate statistical techniques to construct more accurate face models. Many 2D/3D linear/non-linear models =-=[26,39,46,52]-=- have been proposed based on support vector machines, mixture of factor analyzers, Independent Component Analysis, Kernel PCA, etc. See [12,26,52] for an extended review in the context of recognition ... |

90 |
Dynamic Vision: From Images to Face Recognition. 1860941818
- Gong, McKenna, et al.
- 2000
(Show Context)
Citation Context ...deo-conferencing, realistic avatar animation and eye tracking. 1 Introduction Many computer vision researchers have used Principal Component Analysis (PCA) to parameterize appearance, shape or motion =-=[3,10,23,30]-=-. However, one major drawback of this traditional technique is that it needs normalized samples in the training data. In the case of computer vision applications, the result is that the samples have t... |

73 | Transformation-invariant clustering using the EM algorithm. - Frey, Jojic - 2003 |

72 | Robustly estimating changes in image appearance,”
- Black, Fleet, et al.
- 2000
(Show Context)
Citation Context |

57 |
Robust regression
- Li
- 1985
(Show Context)
Citation Context ... conveniently) we can reformulate the minimization problem as one of iteratively reweighted least-squares (IRLS), which provides an approximate, iterative, solution to the robust M-estimation problem =-=[28]. For a giv-=-en σ, a matrix W ∈ ℜd×N , which contains the positive weights for each pixel and each image, is calculated for each iteration as a function of the previous residuals epi = dpt − (πl �k p j=... |

57 |
Robust Statistics - The Approach based on In Functions
- Hampel, Ronchetti, et al.
- 1986
(Show Context)
Citation Context ... 1 (x p ; a l t )). Each element, w pi (p th pixel of the i th image) ofW will be equal to w pi =s(e pi ; p )=e pi , wheres(e pi ; p ) = @(e pi ; p ) @e pi = 2e pi 2 p (e 2 pi + 2 p ) 2 , 10 =-=[28]-=-. Given an initial error, the weight matrix W is computed and Equation (2) becomes: E wereg (B; C;A;)= T X t=1 jjd t L X l=1 ( l Æ ~ B l c t l )(f 1 (x; a l t ))jj 2 W t (3) = T X t=1 L X l=1 jjd ... |

55 | Fast object recognition in noisy images using simulated annealing.
- Betke, Makris
- 1995
(Show Context)
Citation Context ...uences that we tried, the face can move more than 20 pixels from frame to frame). In order to cope with such real conditions, we explore the use of stochastic methods such as Simulated Annealing (SA) =-=[2]-=-, Genetic Algorithms (GA) [27,29] or Condensation (particle filtering) [6,17] for motion estimation. Although the techniques are very similar computationally speaking, here we make use of GA [29] with... |

53 | A D. A framework for automatic landmark identification using a new method of nonrigid correspondence[J
- Hill, Taylor, et al.
(Show Context)
Citation Context ... the image sequence and in some cases the features or curves are placed by hand. If this is the case, the problem is how to put the features in correspondence using rigid or non-rigid transformations =-=[9,29]-=-. In the other direction, Walker et al. [32] have proposed a method for automatically placing landmarks to dene correspondence between images and hence automatically constructing appearance models. S... |

51 | Principal manifolds and bayesian subspaces for visual recognition
- Moghaddam
- 1999
(Show Context)
Citation Context ...ring the last few years there has been a growing trend to apply new machine learning or multivariate statistical techniques to construct more accurate face models. Many 2D/3D linear/non-linear models =-=[26,39,46,52]-=- have been proposed based on support vector machines, mixture of factor analyzers, Independent Component Analysis, Kernel PCA, etc. See [12,26,52] for an extended review in the context of recognition ... |

35 | Early Visual Learning
- Nayar, Poggio
- 1996
(Show Context)
Citation Context ...he samples have to be aligned or geometrically normalized (we assume that other normalizations, e.g. photometric, have already been done). Previous methods for constructing appearance or shape models =-=[10, 16,17,23,30,31]-=- have cropped the region of interest by hand, or have used a hand-labeled pre-defined feature points to apply the translation, scaling and rotation that brought each image into alignment with a protot... |

34 |
Computer Vision for HumanMachine Interaction
- Cipolla, Pentland
- 1998
(Show Context)
Citation Context ...tic. In this paper we focus on the application of face modeling. Most of the previous work on face tracking and modeling is focused on generic trackers, which are independent of the person’s identit=-=y [5,9,10,23,24]. -=-In particular, appearance based face trackers [10,26,sRobust Parameterized Component Analysis 655 34] make use of PCA in order to construct a linear model of the face’s subspace (variation across pe... |

32 | Mixture of eigenfeatures for real-time structure from texture
- JEBARA, RUSSEL, et al.
- 1998
(Show Context)
Citation Context ...e while aligning the training images w.r.t. this subspace. That is, the algorithm that we propose in this paper will simultaneously learn the local appearance basis, creating modular eigenspaces (ME) =-=[26,30]-=- while computing the motion to align the images w.r.t. the ME. The masks which define the spatial domain of the ME are defined by hand in the first frame (no appearance model is previously learned) an... |

32 | Improving identification performance by integrating evidence from sequences - Edwards, Taylor, et al. - 1999 |

30 |
Eigentracking: Robust matching and tracking of objects using view-based representation.
- Black, Jepson
- 1998
(Show Context)
Citation Context ...he construction of facial models using linear subspaces [34]. During recognition or tracking it is common to automatically align the input images with the eigenspace using some optimization technique =-=[4, 10,34]-=-. In contrast, little work has addressed problems posed by facial misalignment at the learning stage. Mis-registration introduces significant non-linearities in the manifold of faces and can reduce th... |

27 |
Separation of texture and shape in images of faces for image coding and synthesis
- Vetter, Troje
- 1997
(Show Context)
Citation Context ...k on face tracking focuses on generic trackers which are independent of the identity of the person being tracked [5,8,10,11,26,27,34], here we focus on Person Specic Facial Appearance Models (PSFAM) =-=[17,26,22,50]-=- for tracking a single individual and use PCA to model the variations due to changes in expression. Although PSFAM are only valid for one person, they remain useful in many vision related applications... |

25 | Dynamic coupled component analysis
- TORRE, D, et al.
- 2001
(Show Context)
Citation Context ...iss666 F. De la Torre and M.J. Black a. b. a. b. a. b. a. b. Fig. 7. a) Original image sequence. b) Normalized face. the result of animating the face with Asymmetric Coupled Component Analysis (ACCA) =-=[13]-=- plus the affine motion of the head. As we can observe this approach allows us to model the rich texture present on the face providing fairly realistic animations. See [13] for further information.sa.... |

24 |
Local feature analysis: a statistical theory for information representation and transmission,
- Penev
- 1998
(Show Context)
Citation Context ...to occlusions [37], etc.). However, it is worth pointing out that representations other than ME have been explored successfully for face recognition and tracking; for instance, Local Feature Analysis =-=[43,42]-=- or Gabor jets with elastic graph matching [51]. Although these techniques have shown good performance in recognition and tracking domains, they do not address the issue of learning a model invariant ... |

20 | Transformation-invariant clustering and dimensionality reduction using em
- Frey, Jojic
- 2000
(Show Context)
Citation Context ...invariant code for natural images. Although he suggests updating the appearance basis, the experiments show only translation-invariant recognition, as proposed by Black and Jepson [4]. Frey and Jojic =-=[21]-=- took a different approach and they introduce an Expectation Maximization (EM) algorithm for factor analysis (similar to PCA) that is invariant to geometric transformations. While their work represent... |

20 |
The approximation of one matrix by another of lower rank
- Eckert, Young
(Show Context)
Citation Context ...n be achieved bysnding the k largest eigenvectors of the covariance matrix DD T [20], here it is useful to exploit work that formulates PCA/Subspace learning as the minimization of an energy function =-=[16,20,21]-=-: E pca (B;C) = jjDBCjj 2 F = T X i=1 jjd i Bc i jj 2 2 = T X t=1 d X p=1 (d pt k X j=1 b pj c jt ) 2 where C = [c 1 c 2 c n ] and each c i is a vector of coeÆcients used to reconstruct the ... |

14 |
Eigenfiltering for flexible eigentracking
- Torre, Vitrià, et al.
- 2000
(Show Context)
Citation Context ...he samples have to be aligned or geometrically normalized (we assume that other normalizations, e.g. photometric, have already been done). Previous methods for constructing appearance or shape models =-=[10, 16,17,23,30,31]-=- have cropped the region of interest by hand, or have used a hand-labeled pre-defined feature points to apply the translation, scaling and rotation that brought each image into alignment with a protot... |

14 | Development of localized oriented receptive fields by learning a translationinvariant code for natural images
- Rao, Ballard
- 1998
(Show Context)
Citation Context ...motions remains unclear. As Schweitzer notices [33] the algorithm is likely to get stuck in local minima, since it comes from a linearization and uses gradient descent methods. On the other hand, Rao =-=[32]-=- has proposed a neural-network which can learn a translation-invariant code for natural images. Although he suggests updating the appearance basis, the experiments show only translation-invariant reco... |

13 |
A probabilisitc framework for rigid and non-rigid appearance based tracking and recognition
- Torre, Yacoob, et al.
- 2000
(Show Context)
Citation Context ...he samples have to be aligned or geometrically normalized (we assume that other normalizations, e.g. photometric, have already been done). Previous methods for constructing appearance or shape models =-=[10, 16,17,23,30,31]-=- have cropped the region of interest by hand, or have used a hand-labeled pre-defined feature points to apply the translation, scaling and rotation that brought each image into alignment with a protot... |

12 | Model-based estimation of facial expression parameters from image sequences
- Eisert, Girod
- 1997
(Show Context)
Citation Context ...te one face given another using PSFAMs. In general it is hard to model and animate faces and often complex models encoding the underlying physical musculature of the face are used (e.g. Candide model =-=[23]-=-). Here we learn the PSFAM of two people with parameterized component analysis introduced in this paper. Then, we manually select all pairs of corresponding images which share a common emotional state... |

11 | Hierarchical model- based motion estimation - Bergen, Anandan, et al. - 1992 |

10 | Optimal eigenfeature selection by optimal image representation
- Schweitzer
- 1999
(Show Context)
Citation Context ...available in the eigenspace, the motion parameters are calculated in a robust manner. However, the method assumes that an initial eigenspace is learned from a training set aligned by hand. Schweitzer =-=[33]-=- has proposed a deterministic method which registers the images with respect to their eigenfeatures, applying it to the flower garden sequence for indexing purposes. However, the assumption of affine ... |

10 | Facial Feature Tracking and Pose Estimation in Video Sequences by Factorial Coding of the Low-Dimensional Entropy Manifolds due to the Partial Symmetrie s of Faces
- Mandel, Penev
- 2000
(Show Context)
Citation Context ...of possible spatial transformations, and can be too computationally intensive when working with realistic high dimensional (greater than two) motion models. Using a dierent approach Mandel and Penev =-=[36]-=- report the interesting observation that non-properly aligned data lie on curved manifolds. This observation forms the basis of an algorithm to align visual data. Results where reported on image seque... |

9 |
A and Pentland A P. Eigenfaces for recognition.Journal of Cognitive Neuroscience
- Turk
- 1991
(Show Context)
Citation Context ... solving a linear system of equations. 2.2 Adding Motion into the Subspace Formulation Principal component analysis has been widely applied to the construction of facial models using linear subspaces =-=[34]-=-. During recognition or tracking it is common to automatically align the input images with the eigenspace using some optimization technique [4, 10,34]. In contrast, little work has addressed problems ... |

9 |
Principal Component Neural Networks (Therory and Applications
- Diamantaras
- 1996
(Show Context)
Citation Context ...the subspace of maximum variation of D 1 . Although a closed form solution for computing the principal components (B) can be achieved bysnding the k largest eigenvectors of the covariance matrix DD T =-=[20]-=-, here it is useful to exploit work that formulates PCA/Subspace learning as the minimization of an energy function [16,20,21]: E pca (B;C) = jjDBCjj 2 F = T X i=1 jjd i Bc i jj 2 2 = T X t=1 d X p=... |

6 |
Automatic learning of appearance face models
- Torre
- 2001
(Show Context)
Citation Context ... we will apply Robust Parameterized Component Analysis to learn a PSFAM and will illustrate the method with applications involving facial modeling. Preliminary results of this paper were presented in =-=[12]-=-. 2 Previous Work This paper is related to previous work on subspace learning methods and PCA. It is beyond the scope of the paper to review all possible applications of PCA, therefore we just briefly... |

6 |
Statistical models of face images—Improving specificity
- Edwards, Lanitis, et al.
- 1998
(Show Context)
Citation Context ...nimation (to animate faces from audio), facial animation in general, video-conferencing, face verification, etc, which usually involve a particular user. In related but different work, Edwards et al. =-=[19,20]-=- have proposed a method for approximately isolating the sources of image variation such as identity, pose, lighting, etc [19] by using linear discriminant analysis. Edwards et al. [20] use this factor... |

6 |
Determining correspondences for statistical models of appearance
- Walker, Taylor
- 2000
(Show Context)
Citation Context ...ures or curves are placed by hand. If this is the case, the problem is how to put the features in correspondence using rigid or non-rigid transformations [9,29]. In the other direction, Walker et al. =-=[32]-=- have proposed a method for automatically placing landmarks to dene correspondence between images and hence automatically constructing appearance models. See the report of Cootes and Taylor [12] for ... |

5 | View alignment with dynamically updated affine tracking - Torre, Gong, et al. - 1998 |

4 | Locating facial features using genetics algorithms
- Lanitis, Hill, et al.
- 1995
(Show Context)
Citation Context ...e can move more than 20 pixels from frame to frame). In order to cope with such real conditions, we explore the use of stochastic methods such as Simulated Annealing (SA) [2], Genetic Algorithms (GA) =-=[27,29]-=- or Condensation (particle filtering) [6,17] for motion estimation. Although the techniques are very similar computationally speaking, here we make use of GA [29] within a coarse-to-fine strategy. Giv... |

3 | la Torre, “Automatic learning of appearance face models - de |

3 |
Development of localized oriented receptive by learning a translation-invariant code for natural images
- Rao
- 1998
(Show Context)
Citation Context ...era motions remains unclear. As Schweitzer notices [47] the algorithm is likely to get stuck in local minima, since it comes from a linearization and uses gradient descent methods. Alternatively, Rao =-=[45]-=- has proposed a neural-network which can learn a translation-invariant code for natural images. Although he suggests updating the appearance basis, the experiments show only translation-invariant reco... |

2 |
Improving identitification performance by integrating evidence from sequences
- Edwards, Taylor, et al.
- 1999
(Show Context)
Citation Context ...CA in order to construct a linear model of the face’s subspace (variation across people) rather than the intra-person variations due to changes in expression. When working with person-specific model=-=s [15,17,20,23], -=-PCA will model the complex intra-person appearance changes due mostly to variations of expression (eyes’ blinking, wrinkles in the mouth area, appearance of the teeth, etc.) rather than modeling the... |

2 |
A framework for robust subspace learning. Accepted for publication
- Torre, Black
- 2001
(Show Context)
Citation Context ...s a parameter that controls the convexity of the robust function and is used for deterministic annealing [4,7]. Benets of the robust formulation for subspace related problems are explained elsewhere =-=[15,16]-=-. Observe that the previous equation is a patched version of Eigentracking [4], and similar to AAM [11] or Flexible Eigentracking [18] without shape constraints. However, in contrast to these approach... |

1 |
Diamantaras.Principal Component Neural Networks (Therory and Applications
- I
- 1996
(Show Context)
Citation Context ...ce of maximum variation of the data D. Although a closed form solution for computing the principal components (B) can be achieved by computing the k largest eigenvectors of the covariance matrix DD T =-=[18]-=-, here it is useful to exploit work that formulates PCA as the minimization of an energy function [14,18]. Related formulations have been studied in various communities 1 Bold capital letters denote a... |

1 |
Fastj reliable tracking under varying illumination
- Casia, Sclaroff
- 1999
(Show Context)
Citation Context ...ating a subspace for the appearance variation. 2.3 Person specic models While most work on face tracking focuses on generic trackers which are independent of the identity of the person being tracked =-=[5,8,10,11,26,27,34]-=-, here we focus on Person Specic Facial Appearance Models (PSFAM) [17,26,22,50] for tracking a single individual and use PCA to model the variations due to changes in expression. Although PSFAM are o... |

1 |
Optimal eigenfeature selection by optimal image registration
- Schewitzer
- 1999
(Show Context)
Citation Context ...available in the eigenspace, the motion parameters are calculated in a robust manner. However, the method assumes that an initial eigenspace is learned from a training set aligned by hand. Schweitzer =-=[47]-=- has proposed a deterministic method which registers a set of images with respect to their eigenfeatures, applying it to thesower garden sequence for indexing purposes. However, the assumption of aÆne... |

1 |
von der Malsburg, Face recognition using by elastic bunch graph matching
- Wiskott, Fellous, et al.
- 1997
(Show Context)
Citation Context ...inting out that representations other than ME have been explored successfully for face recognition and tracking; for instance, Local Feature Analysis [43,42] or Gabor jets with elastic graph matching =-=[51]-=-. Although these techniques have shown good performance in recognition and tracking domains, they do not address the issue of learning a model invariant to geometric transformations. 3 Generative mode... |

1 |
View alignment with dynamically updated aÆne tracking
- Torre, Gong, et al.
- 1998
(Show Context)
Citation Context ...e of interesting points through an image sequence is known, learning the appearance model is straightforward, and if the appearance is known solving for the correspondence is easy. De la Torre et al. =-=[17]-=- proposed a method for face tracking which recovers aÆne parameters using subspace methods. This method dynamically updates the eigenspace by utilizing the most recent history. The updating algorithm ... |

1 |
Eigen for eigentracking
- Torre, Vitria, et al.
- 2000
(Show Context)
Citation Context ...spect to the face model is a common step as noted by Martinez [37], little work has addressed problems posed by misalignment at the learning stage. Previous methods for constructing appearance models =-=[11,18,19,26,40,41]-=- have cropped the region of interest by hand, or have used a hand-labeled, predened, feature points to compute the translation, scaling and rotation that brought each image into alignment with a prot... |

1 |
Improving identiti performance by integrating evidence from sequences
- Edwards, Taylor, et al.
- 1999
(Show Context)
Citation Context ...k on face tracking focuses on generic trackers which are independent of the identity of the person being tracked [5,8,10,11,26,27,34], here we focus on Person Specic Facial Appearance Models (PSFAM) =-=[17,26,22,50]-=- for tracking a single individual and use PCA to model the variations due to changes in expression. Although PSFAM are only valid for one person, they remain useful in many vision related applications... |

1 |
Recognizing imprecisely localized, partially occluded and expression variant faces from a single sample per class. Accepted for publication
- Mart
(Show Context)
Citation Context ...ue is that it requires normalized (aligned) samples in the training data. While, in the recognition process, alignment of the data with respect to the face model is a common step as noted by Martinez =-=[37]-=-, little work has addressed problems posed by misalignment at the learning stage. Previous methods for constructing appearance models [11,18,19,26,40,41] have cropped the region of interest by hand, o... |