## Face identification by fitting a 3D morphable model using linear shape and texture error functions (2002)

### Cached

### Download Links

- [informatik.unibas.ch]
- [mi.informatik.uni-siegen.de]
- [gravis.cs.unibas.ch]
- DBLP

### Other Repositories/Bibliography

Venue: | in European Conference on Computer Vision |

Citations: | 65 - 1 self |

### BibTeX

@INPROCEEDINGS{Romdhani02faceidentification,

author = {Sami Romdhani and Volker Blanz and Thomas Vetter},

title = {Face identification by fitting a 3D morphable model using linear shape and texture error functions},

booktitle = {in European Conference on Computer Vision},

year = {2002},

pages = {3--19}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract This paper presents a novel algorithm aiming at analysis and identification of faces viewed from different poses and illumination conditions. Face analysis from a single image is performed by recovering the shape and textures parameters of a 3D Morphable Model in an analysis-by-synthesis fashion. The shape parameters are computed from a shape error estimated by optical flow and the texture parameters are obtained from a texture error. The algorithm uses linear equations to recover the shape and texture parameters irrespective of pose and lighting conditions of the face image. Identification experiments are reported on more than 5000 images from the publicly available CMU-PIE database which includes faces viewed from 13 different poses and under 22 different illuminations. Extensive identification results are available on our web page for future comparison with novel algorithms. 1

### Citations

3146 |
Multiple View Geometry in Computer Vision
- Hartley, Zisserman
- 2000
(Show Context)
Citation Context ...on-linear interaction of the parameters. Therefore, to recover these parameters, a Levenberg-Marquardt optimisation [18] is used to minimise the geometrical error between s of 2d and the model points =-=[14]-=-: arg min f,φ,γ,θ,t2d � s of 2d − (f P R (s + S α + t3d) + t2d) � 2 = ( ˜ f, ˜ φ, ˜γ, ˜ θ, ˜t2d) (17) The parameters α and t3d are not optimised. The current value of α is used and t3d is uniquely det... |

1860 |
Numerical Recipes in C, The Art of Scientific Computing
- Press
- 1995
(Show Context)
Citation Context ... is straightforward. After subtracting their mean, s, the exemplars are arranged in a data matrix A and the eigenvectors of its covariance matrix C are computed using the Singular Value Decomposition =-=[18]-=- of A: s = 1 �M M i=1 sex i , ai = sex i − s, A = (a1, a2, . . . , aM ) = UΛVT , C = 1 M AAT = 1 M UΛ2 U T The M columns of the orthogonal matrix U are the eigenvectors of the covariance matrix C, and... |

1491 | Active appearance models
- Cootes, Edwards, et al.
- 1998
(Show Context)
Citation Context ...] which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher [12], then Sclaroff and Isidoro [20] and Cootes et al. =-=[8]-=- to derive an optimisation-based algorithm in the specific context of model registration. This new approach, referred to as Image Difference Decomposition (IDD), makes the assumption that there is a l... |

1036 |
Active shape models – their training and application
- Cootes, Taylor, et al.
- 1995
(Show Context)
Citation Context ...mination (see experiments in Section 7) originates from the fact that it is based on a model which closely approximates physical reality. LiST bears some resemblance with the Active Shape Model (ASM) =-=[10]-=- and with the Vectorizer [3]. The ASM recovers the 2D shape of objects, represented as a sparse number of landmark points, using an equation similar to the Equation 18. Apart from the shape representa... |

758 | A morphable model for the synthesis of 3D faces
- Blanz, Vetter
- 1999
(Show Context)
Citation Context ...to find the model parameters which minimise the difference between the image produced by the model and the novel image. The minimisation can be performed by a standard gradient descent algorithm [15] =-=[6]-=- or by a LevenbergMarquardt [17] which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher [12], then Sclaroff and ... |

336 | Lambertian reflectance and linear subspaces
- Basri, Jacobs
- 2003
(Show Context)
Citation Context ...m is to implement it in a coarse to fine fashion. We are planning to evaluate the impact of a coarse to fine implementation on the algorithm performances. Illumination Recovery Recently, Basri et al. =-=[1]-=- and Ramamoorthi et al. [19] have shown that the diffuse part of the illumination was well approximated by a 9D linear space. We want to leverage these findings in our algorithm and recover the light ... |

186 | A signal-processing framework for inverse rendering
- Ramamoorthi, Hanrahan
- 2001
(Show Context)
Citation Context ...oarse to fine fashion. We are planning to evaluate the impact of a coarse to fine implementation on the algorithm performances. Illumination Recovery Recently, Basri et al. [1] and Ramamoorthi et al. =-=[19]-=- have shown that the diffuse part of the illumination was well approximated by a 9D linear space. We want to leverage these findings in our algorithm and recover the light parameters also using linear... |

183 |
Fundamentals of Interactive Computer Graphics
- Foley, Dam
- 1982
(Show Context)
Citation Context ... RφRγRθ of all N vertices. P is the 2N × 3N orthographic projection matrix for the ensemble of vertices. Note that, for rendering, a test of visibility must still be performed using a z-buffer method =-=[11]-=-. Illumination and Colour Transformation We assume that the face is illuminated by ambient light and one directed light. In the same fashion as the shape transformation aforementioned, we denote by ˆt... |

159 | The cmu pose, illumination, and expression (pie) database
- Sim, Baker, et al.
- 2002
(Show Context)
Citation Context ...tification by Fitting a 3D Morphable Model using LiST 11 In this sections, the fitting and identification performances are investigated on the Pose Illumination and Expression (PIE) database from CMU =-=[21]-=- which exhibits combined variations of pose and directed light. None of the 68 individuals in the PIE database is in the training set of 3D scans. We selected two portions of the database to carry out... |

113 |
Image representations for visual learning
- Beymer, Poggio
(Show Context)
Citation Context ...Section 7) originates from the fact that it is based on a model which closely approximates physical reality. LiST bears some resemblance with the Active Shape Model (ASM) [10] and with the Vectorizer =-=[3]-=-. The ASM recovers the 2D shape of objects, represented as a sparse number of landmark points, using an equation similar to the Equation 18. Apart from the shape representation (dense physically-based... |

102 | Face identification across different poses and illuminations with a 3D morphable model
- Blanz, Romdhani, et al.
- 2002
(Show Context)
Citation Context ...he iterative algorithm is stopped when �δt� 2 obtained after this iteration becomes stationary. 6 Discussion Building on our earlier 3D morphable model fitting using stochastic gradient descent (SGD) =-=[5]-=-, we now developed LiST, an algorithm which takes advantage of the linear parts of the model. Applied to the problem of face recognition, we achieved with LiST a similar performance than the SGD algor... |

89 | Active blobs
- Sclaroff, Isidoro
- 1998
(Show Context)
Citation Context ... LevenbergMarquardt [17] which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher [12], then Sclaroff and Isidoro =-=[20]-=- and Cootes et al. [8] to derive an optimisation-based algorithm in the specific context of model registration. This new approach, referred to as Image Difference Decomposition (IDD), makes the assump... |

82 | Resynthesizing facial animation through 3d model-based tracking
- PIGHIN, SZELISKI, et al.
- 1999
(Show Context)
Citation Context ...ich minimise the difference between the image produced by the model and the novel image. The minimisation can be performed by a standard gradient descent algorithm [15] [6] or by a LevenbergMarquardt =-=[17]-=- which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher [12], then Sclaroff and Isidoro [20] and Cootes et al. [... |

74 |
Multidimensional morphable models: a framework for representing and matching object classes
- Jones, Poggio
- 1998
(Show Context)
Citation Context ...l is to find the model parameters which minimise the difference between the image produced by the model and the novel image. The minimisation can be performed by a standard gradient descent algorithm =-=[15]-=- [6] or by a LevenbergMarquardt [17] which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher [12], then Sclaroff ... |

64 | Quo vadis face recognition
- Gross, Shi, et al.
- 2001
(Show Context)
Citation Context ...s); 2 cameras were placed above and below the central camera; and 2 cameras were positioned in the corners of the room. (b) 3D locations of the cameras, the flashes and the head. Courtesy of [21] and =-=[13]-=-. 7.1 Fitting Before fitting a face image, the face must be detected, its pose estimated and the direction of the light must be known. During execution of the algorithm, it refines the estimate of the... |

60 |
Hierarchical motion-based frame rate conversion,” David Sarnoff Res
- Begen, Hingorani
- 1990
(Show Context)
Citation Context ...d by setting the exemplar faces in full correspondence with respect to a reference shape. Correspondences between all exemplar face and the reference face are established by an optical flow algorithm =-=[2]-=-. This produces a 3D deformation field for each exemplar faces which is used to warp the textures onto the reference shape yielding a shape-free texture. This scheme introduces a consistent labelling ... |

55 | Projective registration with difference decomposition
- Gleicher
- 1997
(Show Context)
Citation Context ...scent algorithm [15] [6] or by a LevenbergMarquardt [17] which is computationally expensive as it computes at each iteration the gradient of the image difference. This predicament lead first Gleicher =-=[12]-=-, then Sclaroff and Isidoro [20] and Cootes et al. [8] to derive an optimisation-based algorithm in the specific context of model registration. This new approach, referred to as Image Difference Decom... |

18 |
Face Recognition Vendor Test 2000,” Evaluation Report
- Blackburn, Bone, et al.
- 2001
(Show Context)
Citation Context ...and mouth coordinates)) is in [13] using Visionics FaceIt, one of the most successful commercial face recognition system. FaceIt finished as the top performer in the Face Recognition Vendor Test 2000 =-=[4]-=-. Visionics claim that FaceIt can handle pose variation up to 35 degrees in all direction. Our identification results outperform FaceIt in all 169 cells. For front gallery and probe views, our algorit... |

2 |
face recognition technology) recognition algorithm development and test results
- FERET
- 1993
(Show Context)
Citation Context ...sk of separating different sources of image variation. The recent FERET tests revealed that recognition performance drops significantly when pose and/or directed light are present in the input images =-=[16]-=-. Solving these two problems has become a major research issue. In this paper we present a novel algorithm able to identify faces in presence of combined pose and illumination variation. A common appr... |

1 |
Automatische Rekonstruction der dreidimensionalen Form von Gesichtern aus einem Einzelbild
- Blanz
- 2001
(Show Context)
Citation Context ...ntrast c. The resulting colour of the vertex rendered in the image at the position ˆs2d is ˆtM: ⎛ ⎞ 0.3 0.59 0.11 ˆtM = G M(c) ˆtl + ô, M(c) = c I + (1 − c) ⎝ 0.3 0.59 0.11 ⎠ (10) 0.3 0.59 0.11 j (6) =-=(7)-=-sFace Identification by Fitting a 3D Morphable Model using LiST 5 Again, the dependency of the rendered texture over the model parameters β is clarified by summarising Equations 4, 9 and 10 as one: tM... |