## Estimation of pose and illuminant direction for face processing (1997)

### Cached

### Download Links

- [publications.ai.mit.edu]
- [publications.ai.mit.edu]
- [hera.itc.it:3003]
- DBLP

### Other Repositories/Bibliography

Venue: | Image and Vision Computing |

Citations: | 27 - 1 self |

### BibTeX

@ARTICLE{Brunelli97estimationof,

author = {Roberto Brunelli},

title = {Estimation of pose and illuminant direction for face processing},

journal = {Image and Vision Computing},

year = {1997},

volume = {15},

pages = {741--748}

}

### OpenURL

### Abstract

In this paper three problems related to the analysis of facial images are addressed: the estimation of the illuminant direction, the compensation of illumination effects and, finally, the recovery of the pose of the face, restricted to in-depth rotations. The solutions proposed for these problems rely on the use of computer graphics techniques to provide images of faces under different illumination and pose, starting from a database of frontal views under frontal illumination.

### Citations

1974 | T.: An iterative image registration technique with an application to stereo vision
- Lucas, Kanade
- 1981
(Show Context)
Citation Context ... X by using the correspondence map M in the following way [9]: X(x; y) ! X(x; y) + \Delta(M x (x; y); M y (x; y)) (9) The pixel to pixel correspondence M can be computed using optical flow algorithms =-=[10]-=-, [11], [12], [13]. However, in order to use such algorithms effectively, it is often necessary to pre-adjust the geometry of image X to that of I A [14]. This can be done by locating relevant feature... |

1828 |
Robust Statistics
- Huber
- 1981
(Show Context)
Citation Context ...have zeroaverage and unit variance. The characteristics of this similaritymeasure are extensively discussed in [21] where it is shown that it is less sensitive to noise than xy and technically robust =-=[22]-=-. Hierarchical approaches to the computation of correlation, such as those proposed in [23] are readily extended to the use of this alternative coe cient. The in uence of template shape can be further... |

847 | The design and use of steerable filters
- Freeman, Adelson
- 1991
(Show Context)
Citation Context ...such as those outlined in Appendix B can be used to locate one of the eyes even under large left-right rotations and the direction of the interocular axis can be computed using the method reported in =-=[17]-=-. Let us assume for simplicity of notation that the interocular axis is horizontal. Using the projection techniques reported in [18], [5] we can approximately localize the region were both eyes are co... |

828 |
Computer Vision
- Ballard, Brown
- 1982
(Show Context)
Citation Context ...e features without knowing the identity of the represented person. The usual way to locate a pattern within an image is to search for the maximum of the normalized cross-correlation coefficient ae xy =-=[15]-=-. The sensitivity of this coefficient to changes in the illumination can be reduced by a suitable processing of the images prior to the comparison (see Appendix A). Furthermore, the identity of the pe... |

563 | Face recognition: features versus templates
- Brunelli, Poggio
- 1993
(Show Context)
Citation Context ... author at CBCL-MIT and at IRST. [4] for other approaches). The images for which the direction must be computed are very constrained: they are frontal views of faces with a fixed interocular distance =-=[5]-=-. Once the illuminant direction is known it can be compensated for, obtaining an image under standard illumination which can be more easily compared to a database of faces using standard techniques su... |

136 | Face recognition under varying pose
- Beymer
- 1994
(Show Context)
Citation Context ... can be computed using optical flow algorithms [10], [11], [12], [13]. However, in order to use such algorithms effectively, it is often necessary to pre-adjust the geometry of image X to that of I A =-=[14]-=-. This can be done by locating relevant features of the face, such as the nose and mouth, and warping imagesX so that the location of these features is the same as in the reference image. The algorith... |

127 |
A low-dimensional representation of human faces for arbitrary lighting conditions
- Hallinan
- 1994
(Show Context)
Citation Context ...e addressed problems using learning techniques. 2. Learning the illuminant direction In this section the computation of the direction of the illuminant is considered as a learning task (see [1], [2], =-=[3]-=-, 1 This report describes research done by the author at CBCL-MIT and at IRST. [4] for other approaches). The images for which the direction must be computed are very constrained: they are frontal vie... |

115 |
Deformable templates for face recognition
- Yuille
- 1991
(Show Context)
Citation Context ...roduced: l 0 (y) = 1 n X x 1 ; jF I(x+y) ; (T (x)) T (x)j jF I(x+y)j + j(T (x)) T (x)j (20) The new coe cient introduces the possibility oflocal deformation in the computation of similarity (see also =-=[24]-=- for an alternative approach). References [1] A. P. Pentland. Local Shading Analysis. In From Pixels to Predicates, chapter 3. Ablex Publishing Corporation, 1986. [2] A. Shashua. Illumination and View... |

109 | On photometric issues in 3D visual recognition from a single 2D image
- Shashua
- 1997
(Show Context)
Citation Context ...on In this section the computation of the direction of the illuminant is considered as a learning task (see [1], [2], [3], 1 This report describes research done by the author at CBCL-MIT and at IRST. =-=[4]-=- for other approaches). The images for which the direction must be computed are very constrained: they are frontal views of faces with a fixed interocular distance [5]. Once the illuminant direction i... |

101 | Example based image analysis and synthesis
- Beymer, Shashua, et al.
- 1993
(Show Context)
Citation Context ...rrespondence map M in the following way [9]: X(x; y) ! X(x; y) + \Delta(M x (x; y); M y (x; y)) (9) The pixel to pixel correspondence M can be computed using optical flow algorithms [10], [11], [12], =-=[13]-=-. However, in order to use such algorithms effectively, it is often necessary to pre-adjust the geometry of image X to that of I A [14]. This can be done by locating relevant features of the face, suc... |

88 |
Local shading analysis
- Pentland
- 1984
(Show Context)
Citation Context ...solving the addressed problems using learning techniques. 2. Learning the illuminant direction In this section the computation of the direction of the illuminant is considered as a learning task (see =-=[1]-=-, [2], [3], 1 This report describes research done by the author at CBCL-MIT and at IRST. [4] for other approaches). The images for which the direction must be computed are very constrained: they are f... |

66 |
Smart sensing within a pyramid vision machine
- Burt
- 1988
(Show Context)
Citation Context ...sively discussed in [21] where it is shown that it is less sensitive to noise than xy and technically robust [22]. Hierarchical approaches to the computation of correlation, such as those proposed in =-=[23] a-=-re readily extended to the use of this alternative coe cient. The in uence of template shape can be further reduced by slightly modifying l(x� y). Let us assume that the template T and the correspon... |

60 |
Hierarchical motion-based frame rate conversion,” David Sarnoff Res
- Begen, Hingorani
- 1990
(Show Context)
Citation Context ...the correspondence map M in the following way [9]: X(x; y) ! X(x; y) + \Delta(M x (x; y); M y (x; y)) (9) The pixel to pixel correspondence M can be computed using optical flow algorithms [10], [11], =-=[12]-=-, [13]. However, in order to use such algorithms effectively, it is often necessary to pre-adjust the geometry of image X to that of I A [14]. This can be done by locating relevant features of the fac... |

43 |
Model-based analysis synthesis image coding (MBASIC) system for a person’s face
- Aizawa, Harashima, et al.
- 1989
(Show Context)
Citation Context ...ems, where an appropriate template can then be chosen to speed up the recognition process [14], and for model based coding systems, such as those which could be used for teleconferencing applications =-=[16]-=-. The idea underlying the proposed algorithm for pose estimation is that of quantifying the asymmetry between the aspect of the two eyes due to in-depth rotation and mapping the resulting value to the... |

31 | A novel approach to graphics
- Poggio, Brunelli
- 1992
(Show Context)
Citation Context ...n illumination between I A and I(`; OE; A): \Delta(x; y) = I A (x; y) \Gamma I(`; OE; A; x; y) (8) 5. apply the transformation \Delta to image X by using the correspondence map M in the following way =-=[9]-=-: X(x; y) ! X(x; y) + \Delta(M x (x; y); M y (x; y)) (9) The pixel to pixel correspondence M can be computed using optical flow algorithms [10], [11], [12], [13]. However, in order to use such algorit... |

23 | Illumination and view position in 3d visual recognition
- Shashua
- 1992
(Show Context)
Citation Context ...ng the addressed problems using learning techniques. 2. Learning the illuminant direction In this section the computation of the direction of the illuminant is considered as a learning task (see [1], =-=[2]-=-, [3], 1 This report describes research done by the author at CBCL-MIT and at IRST. [4] for other approaches). The images for which the direction must be computed are very constrained: they are fronta... |

12 |
The quadric reference surface: Applications in registering views of complex 3d objects
- Shashua, Toelg
- 1994
(Show Context)
Citation Context ...ained by rotating the same 3D model used for the generation of the illumination examples: texture mapping techniques are then used to project a frontal view of a face onto the rotated head (see [13], =-=[19]-=- for alternative approaches to the estimation of pose and synthesis of non frontal views). The figure clearly shows that the asymmetry of the two peaks increases with the amount of rotation. The asymm... |

11 |
Computationally Efficient Motion Estimation Algorithm
- Bergen, Adelson
- 1987
(Show Context)
Citation Context ...using the correspondence map M in the following way [9]: X(x; y) ! X(x; y) + \Delta(M x (x; y); M y (x; y)) (9) The pixel to pixel correspondence M can be computed using optical flow algorithms [10], =-=[11]-=-, [12], [13]. However, in order to use such algorithms effectively, it is often necessary to pre-adjust the geometry of image X to that of I A [14]. This can be done by locating relevant features of t... |

9 |
Strategies of robust object recognition for identification of human faces
- Bichsel
- 1991
(Show Context)
Citation Context ...nd low local contrast and is consequently less sensitive to noise. Recently some claims have been made that the gradient direction eld has good properties of invariance toschanges in the illumination =-=[20]-=-. In the case of the direction eld, where a vector is associated to each single pixel of the image, the similarity can be computed by measuring the alignment of the gradient vectors at each pixel. Let... |

9 | Stochastic Minimization with Adaptive Memory
- Brunelli, Tecchiolli
- 1994
(Show Context)
Citation Context ...a diagonal metric is appropriate for this task. A second network is built using the examples f(x `OE ; OE)g `OE . The networks are trained separately using a stochastic algorithm with adaptive memory =-=[8]-=- for the minimization of the global square error of the corresponding outputs: E ` (ff) = X `OE (` \Gamma G ` (x `OE ; ff)) 2 E OE (ff) = X `OE (OE \Gamma G OE (x `OE ; ff)) 2 The error E ` for the di... |

6 |
Robust estimation of correlation: An application to computer vision
- Brunelli, Messelodi
- 1993
(Show Context)
Citation Context ... 0 )=1 , x 0 = y 0 l(x 0 � y 0 )=0 , x 0 = ;y 0 (19) where x 0 and y 0 are normalized to have zeroaverage and unit variance. The characteristics of this similaritymeasure are extensively discussed i=-=n [21]-=- where it is shown that it is less sensitive to noise than xy and technically robust [22]. Hierarchical approaches to the computation of correlation, such as those proposed in [23] are readily extende... |

5 |
Realistic 3-D simulation of shapes and shadows for image processing
- Thirion
- 1991
(Show Context)
Citation Context ...aining an image under standard illumination which can be more easily compared to a database of faces using standard techniques such as cross-correlation. Let us introduce a very simple lighting model =-=[6]-=-: I = (A + Lffi cos !) (1) where I represents the emitted intensity, A is the ambient energy, ffi is 1 if the point is visible from the light source and 0 otherwise, ! is the angle between the inciden... |

4 |
Networks for approximationand learning
- Poggio, Girosi
- 1990
(Show Context)
Citation Context ...correspond to left-right and top-down displacement respectively. From a rather general point of view the problem of learning can be considered as a problem of function reconstruction from sparse data =-=[7]-=-. The points at which the function value is known represent the examples while the function to be reconstructed is the input/output dependence to be learned. If no additional constraints are imposed, ... |

4 |
Similarity Measures in Computer Vision
- Boninsegna, Rossi
- 1994
(Show Context)
Citation Context ...ntrast images for the face recognition task. B. Alternative Template Matching The correlation coefficient is quite sensitive to noise and alternative estimators of pattern similarity may be preferred =-=[21]-=-, [22]. Such measures can be derived from distances other than the Euclidean, such as the L 1 norm defined by: d 1 (x; y) = n X i=1 jx i \Gamma y i j (18) where n is the dimension of the considered ve... |

3 |
Edge Projections for Facial Feature Extraction
- Brunelli
- 1990
(Show Context)
Citation Context ...f the interocular axis can be computed using the method reported in [17]. Let us assume for simplicity of notation that the interocular axis is horizontal. Using the projection techniques reported in =-=[18]-=-, [5] we can approximately localize the region were both eyes are confined. For each pixel in the region the following map is computed: V (x; y) = ae j@ y C(x; y)j if j@ y C(x; y)jsj@ x C(x; y)j 0 oth... |

1 | Applications in Registering Views of Complex 3D Objects - Surface - 1994 |