D.J. Beymer. Pose-Invariant face recognition using real and virtual views. PhD thesis, M.I.T., 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....muscle contraction is computed. ffl Neural network: Another variant of template matching are neural network based approaches to face recognition, for example applying Kohonen self organizing maps [5] To alleviate dependency on the view, multiple views (also called virtual views) can be used [44, 45]. Two approaches are possible: either collect multiple views, or generate virtual views from one template. 2. Feature based Recognition: Kanade [199] first introduced description of faces using features. Since then, many others have further developed the feature based approach to face recognition ....

....an input text [287, 40] ffl Analysis based method: This method has great potential for enhancing graphics systems. It does not require any special markers or other intrusive devices. System robustness can be enhanced by achieving invariance of lighting conditions, head movements, and background [293, 257, 313, 44]. It offers the possibility to extract subtle facial actions with timing information. Temporal information on muscle contraction could be obtained for facial expression [128] and for speech [101] ffl Performance based method: It has the advantage to reproduce animation with the right timing and ....

D.J. Beymer. Pose-Invariant face recognition using real and virtual views. PhD thesis, M.I.T., 1995.

....images considered as bitmaps, on the other hand, basic vector space operations like addition and linear combination are not meaningful. We have argued therefore that a better way to represent images is to associate with each image a shape vector and a texture vector (see for a review [Poggio and Beymer, 1996]) The shape vector of an example image associates to each pixel in the reference image the coordinates of the corresponding point in the example image. The texture vector contains for each pixel in the reference image the color or gray level value for the corresponding pixel in the example ....

....1995] Using pixelwise correspondence [Vetter and Poggio, 1995] and [Beymer and Poggio, 1995] showed that a good approximation of a new face image can be obtained with as few as 50 base faces, suggesting a low dimensionality for both the shape and the texture spaces. As reviewed by [Poggio and Beymer, 1996] correspondence and the resulting vector structure underlie many of the recent view based approaches to recognition and detection either implicitly or explicitly. Certain special object classes (such as cuboids and symmetric objects) can be proved to be exactly linear classes (see [Poggio and ....

[Article contains additional citation context not shown here]

David Beymer. Pose-Invariant Face Recognition Using Real and Virtual Views. PhD thesis, Massachussetts Institute of Technology, 1996.

....each muscle contraction is computed. ffl Neural network: Another variant of template matching are neural network based approaches to face recognition, for example applying Kohonen self organizing maps [5] To alleviate dependency on the view, multiple views (also called virtual views) can be used [37, 38]. Two approaches are possible: either collect multiple views, or generate virtual views from one template. 2. Feature based Recognition: Kanade [160] first introduced description of faces using features. Since then, many others have further developed the feature based approach to face recognition ....

....(adapted from [32] ffl Analysis based method: This method has great potential for enhancing graphics systems. It does not require any special markers or other intrusive devices. System robustness can be enhanced by achieving invariance of lighting conditions, head movements, and background [242, 211, 260, 37]. It offers the possibility to extract subtle facial actions with timing information. Temporal information on muscle contraction could be obtained for facial expression [107] and for speech [85] ffl Performance based method: It has the advantage to reproduce animation with the right timing and ....

D.J. Beymer. Pose-Invariant face recognition using real and virtual views. PhD thesis, M.I.T., 1995.

....of objects. The first was the class of frontal views of human faces. A face Figure 1: The 63 prototype images used to create the model of human faces. The prototype in the top left corner is the reference image which is the average face in this case. database from David Beymer at the MIT AI Lab [3], was used to create a model of the class of faces. For a second face model see [13] The Beymer face database consisted of 63 faces. The second example object class was a set of side views of cars which contained 40 car images. The matching algorithm as described in section 4 was run with the ....

David Beymer. Pose-Invariant Face Recognition Using Real and Virtual Views. PhD thesis, Massachussetts Institute of Technology, 1996.

....images considered as bitmaps, on the other hand, basic vector space operations like addition and linear combination are not meaningful. We have argued therefore that a better way to represent images is to associate with each image a shape vector and a texture vector (see for a review [ Poggio and Beymer, 1996 ] The shape vector of an example image associates to each pixel in the reference image the coordinates of the corresponding point in the example image. The texture vector contains for each pixel in the reference image the color or gray level value for the corresponding pixel in the example ....

....] Using pixelwise correspondence [ Vetter and Poggio, 1995 ] and [ Beymer and Poggio, 1995 ] showed that a good approximation of a new face image can be obtained with as few as 50 base faces, suggesting a low dimensionality for both the shape and the texture spaces. As reviewed by [ Poggio and Beymer, 1996 ] correspondence and the resulting vector structure underlie many of the recent view based approaches to recognition and detection either implicitly or explicitly. Certain special object classes (such as cuboids and symmetric objects) can be proved to be exactly linear classes (see [ Poggio and ....

[Article contains additional citation context not shown here]

David Beymer. Pose-Invariant Face Recognition Using Real and Virtual Views. PhD thesis, Massachussetts Institute of Technology, 1996.

....5. Go to next level in the pyramid 6. Output the parameters 5 Examples and Results The model described in the previous sections was tested on two different classes of objects. The first was the class of frontal views of human faces. A face database from David Beymer, formerly of the MIT AI Lab [6], was used to create a model of the class of faces. For a second face model see [24] The Beymer face database consisting of 62 faces had been set into correspondence by manually specifying a number of corresponding points and then interpolating to get a dense correspondence field. The ....

David Beymer. Pose-Invariant Face Recognition Using Real and Virtual Views. PhD thesis, Massachussetts Institute of Technology, 1996.

....Reconstruction network. The left hand side and right hand side subfigures depict each reconstructed image after the very first iteration and after the relaxation process, respectively. ories (not the inputs themselves) are the only tools available to us we can still consider their PCA components [25]. The computed average memory qmean = 1 n n X i=1 q i (12) was subtracted from each memory and a 16 by 71 matrix, matrix Q, was constructed. The transposed form of the resulting matrix was multiplied by the matrix itself giving rise to the 16 by 16 covariance matrix. The eigenvalues and the ....

D. J. Beymer. Pose-invariant face recognition using real and virtual views. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 1995.

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