## Automatic Construction of Active Appearance Models as an Image Coding Problem (2004)

### Cached

### Download Links

- [www.ri.cmu.edu]
- [www.cs.cmu.edu]
- [www-2.cs.cmu.edu]
- [www-cgi.cs.cmu.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Transactions on Pattern Analysis and Machine Intelligence |

Citations: | 54 - 3 self |

### BibTeX

@ARTICLE{Baker04automaticconstruction,

author = {Simon Baker and Iain Matthews and Jeff Schneider},

title = {Automatic Construction of Active Appearance Models as an Image Coding Problem},

journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},

year = {2004},

volume = {26},

pages = {1380--1384}

}

### Years of Citing Articles

### OpenURL

### Abstract

The automatic construction of Active Appearance Models (AAMs) is usually posed as finding the location of the base mesh vertices in the input training images. In this paper, we re-pose the problem as an energy-minimizing image coding problem and propose an efficient gradientdescent algorithm to solve it.

### Citations

2801 |
Introduction to Statistical Pattern Recognition, 2nd ed
- Fukunaga
- 1991
(Show Context)
Citation Context ...et of images as accurately as possible with a fixed number of parameters (or bits, even). Perhaps the most well-known image coding problem is the one that leads to Principal Components Analysis (PCA) =-=[13]-=-. Even though PCA is optimal (in a certain sense), its coding power is limited. Often a very large number of parameters are needed to code a set of images accurately. The problem with PCA is that the ... |

1580 | Active appearance models
- Cootes, Edwards, et al.
- 2001
(Show Context)
Citation Context ...tive Appearance Models, automatic construction, unsupervised learning, image coding, inverse compositional image alignment, quadratic smoothness priors.s1 Introduction Active Appearance Models (AAMs) =-=[5]-=-, and the closely related concepts of Active Blobs [21] and 3D Morphable Models (3DMMs) [3, 15, 22], are non-linear, generative models of a certain visual phenomenon. The most frequent application of ... |

987 | Visual learning and recognition of 3-d objects from appearance - Murase, Nayar - 1995 |

802 | A morphable model for the synthesis of 3-D faces
- Blanz, Vetter
- 1999
(Show Context)
Citation Context ...e compositional image alignment, quadratic smoothness priors.s1 Introduction Active Appearance Models (AAMs) [5], and the closely related concepts of Active Blobs [21] and 3D Morphable Models (3DMMs) =-=[3, 15, 22]-=-, are non-linear, generative models of a certain visual phenomenon. The most frequent application of AAMs to date has been face modeling [17], however AAMs and 3DMMs may be useful for modeling other p... |

594 | Hierarchical model-based motion estimation
- Bergen, Anandan, et al.
- 1992
(Show Context)
Citation Context ...to help avoid local minima in image alignment tasks such as those in Equations (14) and (15). Typical examples include processing on a Gaussian pyramid and using progressive transformation complexity =-=[2]. As well as using these heuristics, we add one more to Equation (14). Instead o-=-f optimizing Equation (14) we actually optimizes¢ � §�©��s¢¤£¦¥ � ���¡ £¢ ¥ ¤ ¡ ¢ ¤©¤ ¤�¦��s¨�¨ � �� ����¤§¦©¨�� � � ���... |

499 | Lucas-kanade 20 years on: A unifying framework
- Baker, Matthews
- 2004
(Show Context)
Citation Context ...revious estimate of the shape vectorss� to allow the mesh to move freely in that space.) Note that to minimize the quantity in Equation (16) our AAM fitting algorithm [18] has to be modified slightl=-=y [1]-=-. 3.5 Experimental Results In Figure 1 we present the results of running our algorithm on a set of 100 randomly generated square patterns. Each pattern consists of 4 equally sized squares arranged to ... |

323 | Active appearance models revisited
- Matthews, Baker
- 2004
(Show Context)
Citation Context ...ing power of the optimal Linear Appearance Model can be very weak, however. We now consider coding with AAMs. 3.1 Active Appearance Models There are two components to an Active Appearance Model (AAM) =-=[5, 18]-=-, its shape and its appearance. We first define each component in turn and then how to generate a model instance. (We follow the notation of [18], and as in that paper, do not perform a joint shape/ap... |

205 | Linear object classes and image synthesis from a single example image
- Vetter, Poggio
- 2002
(Show Context)
Citation Context ...e compositional image alignment, quadratic smoothness priors.s1 Introduction Active Appearance Models (AAMs) [5], and the closely related concepts of Active Blobs [21] and 3D Morphable Models (3DMMs) =-=[3, 15, 22]-=-, are non-linear, generative models of a certain visual phenomenon. The most frequent application of AAMs to date has been face modeling [17], however AAMs and 3DMMs may be useful for modeling other p... |

197 | The Helmholtz machine
- Dayan, Hinton, et al.
- 1995
(Show Context)
Citation Context ...nse, it is a combination of the AAM modeling work of [5] and the unsupervised learning work of [14]. Our algorithm is also closely related to the unsupervised learning of generative, graphical models =-=[8,9]. 2 Background: Linear Coding and PCA Suppose we are given a -=-set ofsexample training images from the set to be coded: ¡£¢¥¤§¦©¨ where � ��� ������������������� ¤�� and where ¦ these images as accurate... |

192 | Cootes, Automatic interpretation and coding of face images using flexible models
- Lanitis, Taylor, et al.
- 1997
(Show Context)
Citation Context ...f Active Blobs [21] and 3D Morphable Models (3DMMs) [3, 15, 22], are non-linear, generative models of a certain visual phenomenon. The most frequent application of AAMs to date has been face modeling =-=[17]-=-, however AAMs and 3DMMs may be useful for modeling other phenomena as well [15, 21]. AAMs are normally constructed from training images in which the location of the AAM shape mesh vertices are hand-m... |

184 | A Minimum Description Length Approach to Statistical Shape Modeling. 14th IPMI
- Davies, Cootes, et al.
- 2001
(Show Context)
Citation Context ...ng problem. In this respect, the closest related work is that of Davies et al. who posed the problem of automatically constructing a statistical shape model (as opposed to an AAM) as a coding problem =-=[6, 7]-=-. Image coding is the task of representing a set of images as accurately as possible with a fixed number of parameters (or bits, even). Perhaps the most well-known image coding problem is the one that... |

120 | Automatic construction of 3d statistical deformation models of the brain using nonrigid registration
- Rueckert, Frangi, et al.
(Show Context)
Citation Context ... pre-compute the location of the model mesh in the training images. One such technique is optical flow [4, 15]. A very closely related approach is to non-rigidly align the training data and the model =-=[20]-=-. Finally, feature-point tracking has also been used [23]. All of these approaches are pragmatic, and lack elegance. Ultimately, they use an ad-hoc motion estimation technique to estimate the training... |

112 | Morphable 3d models from video
- Brand
- 2001
(Show Context)
Citation Context ...d a great deal of attention. The most common approach is to use some form of motion estimation to pre-compute the location of the model mesh in the training images. One such technique is optical flow =-=[4, 15]-=-. A very closely related approach is to non-rigidly align the training data and the model [20]. Finally, feature-point tracking has also been used [23]. All of these approaches are pragmatic, and lack... |

92 | Active blobs
- Sclaroff, Isidoro
- 1998
(Show Context)
Citation Context ...rvised learning, image coding, inverse compositional image alignment, quadratic smoothness priors.s1 Introduction Active Appearance Models (AAMs) [5], and the closely related concepts of Active Blobs =-=[21]-=- and 3D Morphable Models (3DMMs) [3, 15, 22], are non-linear, generative models of a certain visual phenomenon. The most frequent application of AAMs to date has been face modeling [17], however AAMs ... |

76 |
Multidimensional morphable models: A framework for representing and matching object classes
- Jones, Poggio
- 1998
(Show Context)
Citation Context ...e compositional image alignment, quadratic smoothness priors.s1 Introduction Active Appearance Models (AAMs) [5], and the closely related concepts of Active Blobs [21] and 3D Morphable Models (3DMMs) =-=[3, 15, 22]-=-, are non-linear, generative models of a certain visual phenomenon. The most frequent application of AAMs to date has been face modeling [17], however AAMs and 3DMMs may be useful for modeling other p... |

57 | Estimating mixture models of images and inferring spatial transformations using the EM algorithm
- Frey, Jojic
- 1999
(Show Context)
Citation Context ... unsupervised learning literature where a number of approaches have recently been proposed that simultaneously solve for an appearance model and the parameters of a transformation relating the images =-=[10, 11, 16]-=-. Most of these approaches, however, assume that the set transformations, or shape component, is given a priori; i.e. it is independent of the training data rather than being learnt from the training ... |

48 |
Transformed component analysis: Joint estimation of spatial transformations and image components,” presented at the
- Frey, Jojic
- 1999
(Show Context)
Citation Context ... unsupervised learning literature where a number of approaches have recently been proposed that simultaneously solve for an appearance model and the parameters of a transformation relating the images =-=[10, 11, 16]-=-. Most of these approaches, however, assume that the set transformations, or shape component, is given a priori; i.e. it is independent of the training data rather than being learnt from the training ... |

26 |
Shape discrimination in the hippocampus using an mdl model
- Davies, Twining, et al.
- 2003
(Show Context)
Citation Context ...ng problem. In this respect, the closest related work is that of Davies et al. who posed the problem of automatically constructing a statistical shape model (as opposed to an AAM) as a coding problem =-=[6, 7]-=-. Image coding is the task of representing a set of images as accurately as possible with a fixed number of parameters (or bits, even). Perhaps the most well-known image coding problem is the one that... |

19 | Automatically Building Appearance Models from Image Sequences using Salient Features
- Walker, Cootes, et al.
- 1999
(Show Context)
Citation Context ...ng images. One such technique is optical flow [4, 15]. A very closely related approach is to non-rigidly align the training data and the model [20]. Finally, feature-point tracking has also been used =-=[23]-=-. All of these approaches are pragmatic, and lack elegance. Ultimately, they use an ad-hoc motion estimation technique to estimate the training data that would have been marked by hand. Hence, they ar... |

11 | Separating appearance from deformation
- Jojic, Simard, et al.
- 2001
(Show Context)
Citation Context ...ata. When constructing an AAM, however, the shape basis is unknown and must be solved for. More recently there has been some work on simultaneously learning both a shape model and an appearance model =-=[12,14]-=-. This work is perhaps the most closely related work to ours in the unsupervised learning literature. Our algorithm can be regarded as solving a similar problem to that of [12, 14], but for the more p... |

6 | Efficient stochastic source coding and an application to a Bayesian network source model
- Frey, Hinton
- 1997
(Show Context)
Citation Context ...nse, it is a combination of the AAM modeling work of [5] and the unsupervised learning work of [14]. Our algorithm is also closely related to the unsupervised learning of generative, graphical models =-=[8,9]. 2 Background: Linear Coding and PCA Suppose we are given a -=-set ofsexample training images from the set to be coded: ¡£¢¥¤§¦©¨ where � ��� ������������������� ¤�� and where ¦ these images as accurate... |

5 | Automatic learning of appearance face models - Torre - 2001 |

2 | Product Analysis: Learning to model observations as products of hidden variables
- Frey, Kanna, et al.
- 2001
(Show Context)
Citation Context ...ata. When constructing an AAM, however, the shape basis is unknown and must be solved for. More recently there has been some work on simultaneously learning both a shape model and an appearance model =-=[12,14]-=-. This work is perhaps the most closely related work to ours in the unsupervised learning literature. Our algorithm can be regarded as solving a similar problem to that of [12, 14], but for the more p... |