D. Beymer, "Vectorizing face images by interleaving shape and texture computations," Artif. Intell. Lab., Mass. Inst. Technol., Cambridge, A.I. memo 1537, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....independently. Their system detects canonical points on the face and uses these landmarks to warp faces to a shape free representation prior to implementing an cigcnfacc technique for charac terizing grayscale variations (face texture) Similarly, the face vcctorizcr system of Bcymcr Poggio [2] uses optical flow to obtain a shape representation decoupled from that of texture (in the form of a 2D correspondence field between a given face and a canonical model) However, one of the difficulties with using optical flow for correspondance between two different individuals is that the ....

David Beyruer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

....for a specific pattern, such as the human face, the number of admissible configurations is a only tiny fraction of that. In other words, the intrinsic dimension is much lower than # ## . Both the shape and texture provide important information useful for characterizing the face appearance [1]. Alignment of a given face to a canonical face enables extraction of refined shape and texture parameters in the coordinate system of the canonical face model. It is crucial for high accuracy face recognition and synthesis [13, 8, 9, 3, 11] The active appearance model (AAM) 5] is a powerful ....

D. Beymer. "Vectorizing face images by interleaving shape and texture computations". A. I. Memo 1537, MIT, 1995.

....given image from the feature subspace (DFFS) as a di erence measure between input pattern and the face class. The method was proven ecient for frontal (or other xed pose) faces [1] 2] 3] In order to represent frontal as well as non frontal faces variations, Beymer, Jones, Vetter and Poggio [4] 5][6], Craw[7] and Cootes and Taylor[8] proposed a non linear model, which combines two parametric linear subspace models (one for face appearances, the other for varying pose) by image warping, to represent the distribution of varying pose faces. With the proposed model, one can synthesize a given ....

....combination of frontal face images, where the warping displacement eld is constrained by a subspace model learnt from face images under varying pose. Alignment involves minimizing an appropriate distance measure between the original image and the synthesized image. Existing techniques in [4] 5][6][7] 8] use an iterative procedure to estimate both texture and pose parameters (see Section 2 for de nitions) During each iteration, the synthesized image is compared with the original image using a pixel wise intensity di erence measure. We note that using intensity di erence measure to quantify ....

[Article contains additional citation context not shown here]

David Beymer, \Vectorizing Face Images by Interleaving Shape and Texture Computations", M.I.T., A.I. Memo, No. 1537, Sept. 1995.

....involves the derivation of salient features from the raw input data in order to reduce the amount of data used for classification and simultaneously provide enhanced discriminatory power. Recently shape and texture ( shape free image) have become prominent for encoding face images [6] [2], 23] 13] Shape and texture coding, usually used in conjunction with norm based coding, is a two stage process once the face has been located. Coding starts by annotating the face using important internal and face boundary points. Once these control points are located, they are aligned using ....

....training images are representative of the range of face (class) variations; otherwise, the performance difference between the PCA and MDF spaces is not significant [21] To further improve PCA stand alone methods, both new face representation approaches and new classifiers are emerging. Beymer [2] introduced a vectorized image representation consisting of shape and texture. Vetter and Poggio [23] used such a vectorized face representation for image synthesis from a single example view. Craw, et al. [7] and Lanitis, Taylor and Cootes [13] developed Mahalanobis distance classifiers for face ....

D. Beymer, "Vectorizing face images by interleaving shape and texture computations," A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

....novel approach to match linear models to novel images that can be used for several visual analysis tasks, including recognition. In this paper we develop the approach in detail and show its performance not only on line drawings but also on gray level images. Among papers directly related to ours, [Beymer, 1995] addressed the problem of modeling human faces in order to generate virtual faces to be used in a face recognition system. A virtual view is a synthetically generated image of a face with a novel pose or expression. Beymer modeled texture the same way we do as a linear combination of ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo 1537, MIT, 1995.

....space, and face detection and identification are then carried out in the reduced space. Under the same PCA scheme, shape and texture ( shape free image) coding has recently become prominent for encoding face images in order to encompass more concrete information for face representation [13] [6], 11] Shape and texture coding, usually used in conjunction with norm based coding, is a two stage process once the face has been located. Coding starts by annotating the face using important internal and face boundary points. Once these control points are located, they are aligned using ....

D. Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

.... on the 2D shape of the object in the image (e.g. optic flow [8] photometric approaches model texture variations assuming shape alignment [21] Current work in the area of image based object modeling deals with the shape (2D) and texture (grayscale) components of an image in an independent manner [3, 5]. Our novel representation combines both the spatial (XY) and grayscale (I) components of the image into a 3D surface (or manifold) and then efficiently solves for a dense correspondance map in the XY I space. This manifold matching technique can be viewed as a more general formulation for ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

.... to derive salient features from the raw input data in order to reduce the amount of data used for classification and simultaneously provide enhanced discriminatory power [3] 12] Recently shape and intensity (texture or shape free image) have become prominent for encoding face images [4] [2], 15] 8] Shape and texture coding, usually used in conjunction with norm based coding, is a two stage process once the face has been located. Coding starts by annotating the face using important internal and face boundary points. Once these control points are located, they are aligned using ....

D. Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

....approach based on flexible models, there is the problem of finding the correspondence between model and image. In our implementation we used a general method for finding this correspondence. However, if the class of objects is known in advance, a method specific to this object class could be used [9, 7]. In this case the correspondence field is linearly modeled by a known set of deformations specific to that class of objects. A second problem, specific to our approach is the existence of linear object classes and the completeness of the available examples. This is equivalent to the questions of ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. to appear as A.I. Memo, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

....involves the derivation of salient features from the raw input data in order to reduce the amount of data used for classification and simultaneously provide enhanced discriminatory power. Recently shape and texture ( shape free image) have become prominent for encoding face images [4] [2], 18] 9] Shape and texture coding, usually used in conjunction with norm based coding, is a two stage process once the face has been located. Coding starts by annotating the face using important internal and face boundary points. Once these control points are located, they are aligned using ....

....problem Liu and Wechsler [10] introduced Enhanced FLD Models (EFM) to improve on the generalization capability of the standard FLD based classifiers such as Fisherfaces. To further improve PCA stand alone methods, both new face representation methods and new classifiers are emerging. Beymer [2] introduced a vectorized image representation consisting of shape and texture. Vetter and Poggio [18] used such a vectorized face representation for image synthesis from a single example view. Craw, Costen and Kato [5] and Lanitis, Taylor and Cootes [9] developed a Mahalanobis distance classifier ....

D. Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

....to match linear models to novel images that can be used for several visual analysis tasks, including recognition. In this paper we develop the approach in detail and show its performance not only on line drawings but also on gray level images. Among papers directly related to ours, Beymer [5] addressed the problem of modeling human faces in order to generate virtual faces to be used in a face recognition system. A virtual view is a synthetically generated image of a face with a novel pose or expression. Beymer modeled texture the same way we do as a linear combination of ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo 1537, MIT, 1995.

....phenomenon [13, 6] Furthermore, we seek to unify the shape and texture components of an image in a single compact mathematical framework. Current work in the area of image based object modeling deals with the shape (2D) and texture (grayscale) components of an image in an independant manner [3, 5]. Our novel representation combines both the spatial (XY) and grayscale (I) components of the image into a 3D surface (or manifold) and then efficiently solves for a dense correspondance map in the XY I space. This manifold matching technique can be viewed as a more general formulation for image ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

.... high dimensional data [9] Furthermore, we use a novel representation for d(I1 ; I2 ) which combines both the spatial (XY) and grayscale (I) components of the image in a unified XYI framework (unlike previous approaches which essentially treat the shape and texture components independently, e.g. [3, 4, 7, 2]) Specifically, I1 is modeled as a physicallybased deformable 3D surface (or manifold) in XYI space which deforms in accordance with attractive physical forces exerted by I2 . The dynamics of this system are efficiently solved for using the analytic modes of vibration [10] yielding a 3D ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

....the case of linear object classes. Furthermore, many object transformations such as 3D rotations of a rigid object and changing expression of a face can be approximated by linear transformations, that can be learned from a small number of examples. The same motivation underlies the work of Beymer [2] who describes an alternative approach, also based on a linear combination of prototypes, to vectorize greylevel images. 3 Model based matching using prototypes 3.1 The model We would like the models used for model based matching to be learned from examples as opposed to being hardwired. To ....

....as opposed to black and white line drawings. In this case, in addition to modeling the shape of objects, we also model the texture of objects. We model texture analogously to the way we modeled shape as a linear combination of the grey level values (texture) of the prototype images (see also [2], for an alternative approach to the same problem) A rather general justification of models of shape and texture consisting of linear combinations of prototypical shapes and textures is the following. Under weak assumptions, one can prove that if any network can learn to synthesize shape or ....

David Beymer. Vectorizing Face Images by Interleaving Shape and Texture Computations. To be published as AI Memo, AI Laboratory, MIT 1995.

....independently. Their system detects canonical points on the face and uses these landmarks to warp faces to a shape free representation prior to implementing an eigenface technique for characterizing grayscale variations (face texture) Similarly, the face vectorizer system of Beymer Poggio [2] uses optical flow to obtain a shape representation decoupled from that of texture (in the form of a 2D correspondence field between a given face and a canonical model) However, one of the difficulties with using optical flow for correspondance between two different individuals is that the ....

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

....The underlying intuition is that such a map is easier to learn because the correspondence vectors factor out lighting effects, and also because they undergo reasonably smooth change during motion of the underlying object to be modeled. A helpful, and often important, way (discussed in Beymer [3]) to think about the distinction between images and correspondences is to view correspondence as a way to sample the motion (or shape) of an object between two views, and to view images as a way to sample the object s texture for those views. Synthesizing novel intermediate correspondences is thus ....

David Beymer, "Vectorizing Face Images by Interleaving Shape and Texture Computations", MIT AI Lab memo, No. 1537, September, 1995.

....of feature points of a new image are computed by finding optical flow between the two images. Thus the shape vector of the new image, really a relative shape, is described by a flow or a vector field of correspondences relative to a standard reference shape. Our face vectorizer (see Beymer [9]) which uses optical flow as a subroutine, is also used to automatically compute the vectorized representation. Optical flow matches features in the two frames using the local grey level structure of the images. As opposed to a feature finder, where the semantics of features is determined in ....

....for interperson correspondence when the two people are similar enough in grey level appearance, but this does not happen frequently enough to be useful. Finally, our image vectorizer is a new method for computing pixelwise correspondence between an input and an average face shape y std . Beymer [9] provides the details; here we only set the problem up and sketch the solution. To model grey level face texture, the vectorizer uses a set of N shape free prototypes t p j , the same set as described before for texture representation. Vectorizing an input image i a means simultaneously solving ....

[Article contains additional citation context not shown here]

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

....of 55 people. Another potential basis set is the original example textures themselves. That is, we approximate t a by a linear combination of the n original image textures t p i b t a = P n i=1 fi i t p i : 4) Using this basis set requires computing the pseudoinverse of T off line; see Beymer [4] for the details. 4.2 Run time vectorization In this section we go over the details of the vectorization procedure. The inputs to the vectorizer are an image i a to vectorize and a texture model consisting of N eigenimages e i and mean image t mean . In addition, the vectorizer takes as input ....

....pose specific vectorizers through interpolation. 5 Application: feature finding Our application of the vectorizer has focused on using the correspondences in the shape component. In this section we describe experimental results from using these correspondences to locate facial features. Beymer [4] also discusses using the correspondences to register the features on two arbitrary faces at the same pose. The shape component y std a Gammastd can be sampled to locate a set of feature points in input i a . First, during off line example preparation, the feature points of interest are located ....

[Article contains additional citation context not shown here]

David Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. Memo No. 1537, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1995.

No context found.

D. Beymer, "Vectorizing face images by interleaving shape and texture computations," Artif. Intell. Lab., Mass. Inst. Technol., Cambridge, A.I. memo 1537, 1995.

No context found.

D. Beymer, "Vectorizing face images by interleaving shape and texture computations," A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

No context found.

D. Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

No context found.

D. Beymer. Vectorizing face images by interleaving shape and texture computations. A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

No context found.

D. Beymer, "Vectorizing face images by interleaving shape and texture computations," A.I. memo No. 1537, Artificial Intelligence Laboratory, MIT, 1995.

No context found.

Beymer, D., "Vectorizing Face Images by Interleaving Shape and Texture Computations," MIT Artificial Intelligence Laboratory, A.I. Memo No. 1537, September 1995. 198 198

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC