- Research Article
- Open Access
Adapted Active Appearance Models
EURASIP Journal on Image and Video Processing volume 2009, Article number: 945717 (2009)
Active Appearance Models (AAMs) are able to align efficiently known faces under duress, when face pose and illumination are controlled. We propose Adapted Active Appearance Models to align unknown faces in unknown poses and illuminations. Our proposal is based on the one hand on a specific transformation of the active model texture in an oriented map, which changes the AAM normalization process; on the other hand on the research made in a set of different precomputed models related to the most adapted AAM for an unknown face. Tests on public and private databases show the interest of our approach. It becomes possible to align unknown faces in real-time situations, in which light and pose are not controlled.
All applications related to face analysis and synthesis (Man-Machine Interaction, compression in video communication, augmented reality) need to detect and then to align the user's face. This latest process consists in the precise localization of the eyes, nose, and mouth gravity center. Face detection can now be realized in real time and in a rather efficient manner [1, 2]; the technical bottleneck lies now in the face alignment when it is done in real conditions, which is precisely the object of this paper.
Since such Active Appearance Models (AAMs) as those described in  exist, it is therefore possible to align faces in real time. The AAMs exploit a set of face examples in order to extract a statistical model. To align an unknown face in new image, the models parameters must be tuned, in order to match the analyzed face features in the best possible way. There is no difficulty to align a face featuring the same characteristics (same morphology, illumination, and pose) as those constituting the example data set. Unfortunately, AAMs are less outstanding when illumination, pose, and face type changes. We suggest in this paper a robust Active Appearance Model allowing a real-time implementation. In the next section, we will survey the different techniques, which aim to increase the AAM robustness. We will see that none of them address at the same time the three types of robustness, we are interested in pose, illumination, and identity. It must be pointed out that we do not consider the robustness against occlusion as  does, for example, when a person moves his hand around the face.
After a quick introduction of the Active Appearance Models and their limitations (Section 3), we will present our two main contributions in Section 4.1 in order to improve AAM robustness in illumination, pose, and identity. Experiments will be conducted and discussed in Section 5 before drawing a conclusion, suggesting new research directions in the last section.
2. State of the Art
We propose to classify the methods which lead to an increase of the AAM robustness as follows. The specific types of dedicated robustness are in italic.
Invariant features (illumination)
Canonical representation (illumination)
Parameter space extension
Light modeling (illumination)
3D modeling (pose)
Models number increasing
Supervised classification (pose/expression)
Unsupervised classification (pose/expression)
Learning base specialization
Hierarchical approach (pose/expression)
Identity specification (identity)
Preprocess methods seek to substitute the AAM texture input for a preprocessed image, in order to minimize the influence of illumination. In Invariant features, an image feature invariant, or a less illumination sensitive variation, is used: an image gradient , specific face features like corner detectors for the eyes and mouth , the concatenation of several colors components (H and S from HSV code and image gradient for example) , wavelet networks , or distance map . Except for the last one, those methods all have a serious drawback: by concatenating the different invariant characteristics, they increase the texture size and therefore the algorithm complexity. Steerable filters  can be used to replace texture information and to characterize the region around each landmarks. The evaluation of those filters increases the algorithm complexity but the amount of information to be process by the AAM remains the same if low resolution models (64 64) are used for real-time application. For high resolution models, a wedgelet representation is proposed  to compress the texture. In a Canonical representation, the illumination variations are normalized  or reduced . The shadows also can be evaluated , in order to recover the face 3D model, and then reproduce a texture without any shadow. Those approaches remain uncertain.
Parameter Space Extension methods increase the number of AAM parameters, in order to model the variability introduced in the learning base, which was used to create the face model. In Light modeling, a subspace in the parameter space is learned and built, in order to control the illumination variation. A modeling throughout the Illumination Cone [15, 16] or Light Fields [17, 18] is suggested. The illumination direction can also be estimated through the construction of a learning base of faces, which were acquired under a number of different illuminations, each of them being created by the variation of a single light source position . The illumination variations are then modeled by the principal component analysis embedded in the AAM. All of those methods make the algorithm cumbersome, since the number of parameters needing optimization is increased, and the parameter space is broken up. The optimization, carried on a bigger and noncompact space parameter, is then more difficult to control. In 3D modeling, the face pose variability is transferred from the appearance parameter space to the sub-space which controls the pose (face position and angle). Reference  introduces a new parameter to be optimized, using the pose information associated to each face represented in the learning base. A 3D AAM can also be used either from the shapes and textures acquired from a scanner , or with a frontal and profile face view of each of the learning base face [22–24]. Reference  enriches the 3D AAM's parameters by using the Candide model parameters related to Action Units to deform the mouth and eyebrows. The 3D approach is clearly relevant to increase the AAM robustness related to the pose variability. Nevertheless, as the 3D model becomes more complex, a real-time implementation remains difficult.
Models number increasing methods specify the classes existing in the parameter space of the AAM parameters and define a specific active model in each of those classes. In Supervised classification, the variability type of the learning base is defined and the classes which make up the parameter space are known: the different face views used for the pose variability [26–29] or the different expressions for the expression variability . A huge model containing each submodel specific to each view can be constructed  by concatenating each shape and texture vectors for each view on two large shape and texture vectors. In Unsupervised classification, the classes which constitute the parameter space are found automatically via K-means  or a Gaussian mixture [33, 34]. For each of these methods, active models are numerous. They must be optimized in parallel, in order to decide which one is best suited for the analyzed face. This is not feasible in real time, in our applicative context. One single model can be used in conjunction with Gaussian mixture  to avoid implausible solution during the AAM convergence.
Learning base specialization methods restrict the search space to only one variability (of one face feature or identity). In Hierarchical approach, face features research is divided in two steps: a rough research of face key points and then a refined analysis of face feature by the mean of a specific model for each face feature (eyes, nose, mouth) [36–39]. Like the previous methods, those approaches consist in increasing the number of active models to be optimized in parallel, and then make the alignment system cumbersome. In Identity specification, the database identity variability is removed. Reference  claims that a generic AAM featuring pose, identity, illumination, and expression variability is less efficient than an AAM dedicated to one identity featuring only pose, illumination, and expression variability. Reference  suggests to perform an on-line identity adaptation on an image sequence, by means of a 3D AAM construction, starting from the first image of the face without any expression. This method is not robust since the first image must be perfectly aligned to allow a good 3D AAM modeling.
None of those methods fulfill our constraints, since none of them take into account unknown faces in variable pose and illuminations, at the same time. Let us recall that our main objective is to keep the AAM real-time aspect, while increasing their robustness. Therefore, we started with Invariant features methods related to illumination robustness, in which the AAM texture is pre-processed, and then later suggested a technique (Section 4.1), which does not increase the AAM computation cost. With regard to the robustness associated with pose and identity, and considering the work presented in Identity specification as a start point, we propose to adapt the active model to the analyzed person by means of precomputed AAMs (Section 4.2).
3. Imitation of Active Appearance Models
Active Appearance Models (AAMs) create a joint model of an object's texture and shape from a database comprising different views of the object. The texture inside the shape is normalized in shape (by means of mean shape warping) and in luminance (by means of gray levels mean and variance) and leads to a free shape texture . Two Principal Component Analyses (PCA) are performed on the shapes and textures examples of the learning base
and are the mean shape and mean texture, and are both vectors representing the variations of the orthogonal modes related to shape and texture, respectively. and are both vectors representing shape and texture parameters. We then apply a third PCA on vectors .
is the matrix of the eigenvectors obtained by PCA. is the appearance parameters vector. To each eigenvector is associated an eigenvalue, which indicates the amount of deformation it can generate. In order to reduce the vector dimension, we keep of the model deformation. It is then possible to synthesize an image of the object with the appearance vector .
When we want to align the object in an unknown image , we shift the model defined by the vector relating to a pose vector :
is the rotation of the model in the image plan, is the scale, and and are, respectively, the gravity centre abscissa and ordinate of the model in the analyzed image. We adjust step by step each component of vector , creating then at each iteration a new shape , and a new texture both normalized in shape and luminance, respectively. Let us now consider the texture associated with the region of the image inside the shape . We warp this texture into the mean shape (1) thanks to the warping function (4), and we perform a photometric normalization (5) using the mean and variance evaluated on the warped texture . The residual error between the texture extracted from the image, and the texture generated by the model is then minimized throughout the model parameters tuning, by means of a pre-computed Jacobian, which links the errors to the appearance and pose vectors variations , or by applying classical optimization techniques like simplex  or gradient descent 
with being the number of pixels inside the texture. After a number of iterations, typically one hundred, the error (7) converges to a small value: the model overlaps the object in the image , and produces an estimation of its shape and texture. Those steps are summarized in Algorithm 1
Optimization. Repeat (a) to (e)
From the model, generate a shape and a texture
Retrieve a nonnormalized texture in the image
Normalize to produce :
Warp in the mean shape(4)
Photometric normalize (5)
Evaluate the error (6)
Tune the model parameters
The number of operations, which are processed during the optimization step (see (8)), is evaluated from the number , which is the number of texture pixels, the appearance vector dimension and the points which make up the shape. does not take into account the warping (Algorithm 1 (c).(i)): it is realized on the GPU, and uses of the total processing time (a CPU warping implementation will reduce the process speed by one hundred)
AAM robustness is then linked to the variability introduced in the learning base. The more this one will contain variability, the more the AAM will be able to adapt itself to variable faces. Unfortunately, it is not possible to force a deformable model, created from a learning base and containing a lot of variability, to converge. In fact, the more the learning base will present a large variability, the more the data represented in the parameters space will form different classes; therefore, holes, that is, regions without any data, will appear. Consequently, it is very difficult to force the AAM to converge in this breaking up space. Figure 1 illustrates this problem. The learning base is realized from thirty faces in five different poses. The projection of those examples on the two first appearance parameters shows clearly four clusters, with each of them being specific to a particular pose. Only the frontal faces and those oriented towards the bottom seem to belong to the same cluster. The manifold in this example is clearly broken up; leading thus to a multi-manifold.
4. Proposed Methods
Our two main contributions consist of the Oriented Map Active Appearance (OMAP) Models to give AAM the capacity to align the face in any illumination conditions; the Adapted AAM for pose and identity robustness.
4.1. OM-AAM: Oriented Map Active Appearance Models
Empirical comparisons in face recognition  show that among the Pre-process methods (see Section 2), the uniform or specific histogram transformations are those which lead to the best recognition rates. For that reason, we propose to apply systematically on the images an adaptive histogram equalization from CLAHE . It consists in splitting the image in eight by eight blocks, and in realizing in each block a specific histogram equalization according to a Rayleigh distribution. A specific equalization function is then attached to each block. In order to be able to reject the side-effect related to each blocks, the final result for each pixel is the bilinear interpolation of the equalization functions, associated to the four neighboring blocks of the evaluated pixel
A comparison  between the Viola and Jones face detector  and Froba's one  shows that their relative performances are equivalent when the background is uniform. The first detector is more efficient when faced with a complex background, but is also more difficult to implement. In our application, faces are previously detected and we must align them. The background does not disturb very much the AAM performances.
For that reason, we started with the works of [46, 47]. These explain how to create, with the original image, two images representing the sines and cosines of the detected angle on each pixel, with the work of , which explains how to generate two images with both horizontal and vertical gradients. We propose to simply use the angle on each pixel instead of its gray level. This angle is evaluated on values. In practice we quantify it on eight bits, so . Under a quantification of six bits the results begin to decrease. The new texture is then made out of an image representing the orientation of each pixel, that we call an oriented map. If and represent the horizontal and vertical gradients evaluated on the image , then the oriented map, whose values evolve between and , is estimated in the following manner:
The function is the fourth quadrant inverse tangent. As we can see in Figure 2, when the edges are coded between and , a discontinuity exists in . The roughly vertical edges generate at the same time very low and high levels of information in the oriented map. We observe the effect of this discontinuity on the right face outline (see Figure 3) which flickers between black (high part of the face outline) and white (low part). We propose to realize a mapping (11) from to with the modulo operation, and the absolute value
As we can see in Figure 2, after the mapping process, the edges close to the vertical (orientation angle close to zero, or ) will get a low level of information on an oriented map and those, close to an horizontal position (orientation angle close to or ), will produce a high level of information.
In order to reduce the noise in uniform regions as illustrated in the background of Figure 3(c), we propose to emphasize the signal correlated with the high gradient information region, as it is suggested by  and to use the following nonlinear function :
with being the element by element multiplication. During the modeling, the oriented textures from images will replace the textures usually used by the AAM.
In the segmentation phase, we evaluate the difference between the texture synthesized thanks to the model and the texture analyzed in the image (Figure 3(f)). This texture, in classical AAM, is normalized in luminance and shape at each iteration. The photometric normalization is no longer necessary in our case, since the new texture results in an angle evaluation. When the object is oriented with an angle of , we shift the model with respect to the vector (3) and evaluate a difference between the original image inside the model obtained shape, and the model obtained texture. The difference between those two textures is made in the reference model: a normalized shape with an orientation .
This is not a problem when we deal with gray levels. In our case, since we have replaced the pixel information by the edges orientation, which is evaluated for each pixel, there is no more rotational invariance. As an example, let us consider the ellipse lying in Figure 4 with a pixel on a 45-degree edge. On an oriented map (Figure 4(a)), this pixel in the reference model will have a value of (if the levels range from to ). If we look at the same rotated ellipse of degrees in a test image (Figure 4(b)), the corresponding pixel on the object will have a null value, since the filters used in order to extract the gradients work in the same direction, despite the object orientation. After the warping which takes into account the pose parameter , the texture of the rotated object will have the same value before and after rotation. The corresponding pixel in the model () will be compared to the image's p pixel ().
In order to compare the model texture to that of the object despite its orientation in the image, we simply subtract, before that comparison, an offset (14) to the levels produced by the oriented map. This offset is linked to the pose parameter in the following manner:
We can see in Figure 4(c) that this operation allows the comparison of the orientation information lying in the model texture and the analyzed image texture, whatever the object orientation is.
In order to be able to subtract the offset (14), we need to keep the original values of the edge angle, detected in the image. Therefore, we propose to evaluate, during the segmentation phase, the oriented map between and in the pre-process step (Algorithm 2 ().(b)), and to realize at each iteration, during the optimization phase, the mapping (Algorithm 2 ().(c).(ii)) and the product (Algorithm 2 ().(c).(iii)) operated by the nonlinear function . This function is evaluated during the pre-process (Algorithm 2 ().(c)) and is, then, not time consuming. This new segmentation proposition is summarized by the following Algorithm 2.
Histogram equalization (CLAHE)(9)
Oriented map generation: angle range from to (10)
Evaluate the non-linear function (12)
Optimization. Repeat (a) to (e)
On the basis of the model, generate a shape and a texture
Retrieve a nonnormalized texture in the image
Normalize to produce :
Add the offset angle to the texture(14)
Map the orientation from to (11)
Multiply each pixel by the nonlinear function evaluated in step ().(c)
Warp the new texture in the mean shape to produce
Evaluate the error
Tune the model parameters.
The cost overrun generated by the oriented map is in the order of operations. In real context, we use a texture of pixels and a shape of key points for an appearance vector comprising approximately parameters (see (8)). The optimization cost overrun is , bearing in mind that the warping consumes fifty percent of the process time. In our implementation, we effectively observe a similar increase ( to be precise) when we compare the process time related to the classical AAM, and the one related to our proposition, pre-process step included (Algorithm 2 ()).
As previously said in Section 3.3, the AAM robustness is related to the face variability in the learning base. A great variability induces a multi-manifold parameter space which disturbs the AAM convergence. Instead of using a very generic model containing a lot of variability, we suggest to use an initial model , which contains only a variability in identity, and then use a specific model , containing variability in pose and expression.
4.2.1. Initial Model
Let a general database contain three types of variability: expression, identity, and pose (see Figure 5). We do not include illumination variability in this database since this variability was treated in the preceding sections. It is made of several different faces, holding four distinct expressions: neutral, , , and . Each of the faces presents each of those expressions for the five different poses: frontal face, looking up, left, right, and looking down.
The initial model is realized from a database containing different neutral expression frontal faces (see Figure 6). We use only the images on the horizontal axis of the general database. This initial model will be used to perform a rough alignment on the unknown face.
4.2.2. Type Identification of the Analyzed Face
Let be the appearance vector after the alignment of the model on the unknown analyzed face. In the parameter space of the model parameters, we seek for the nearest parameters vectors of belonging to the learning initial database . Those nearest neighbors correspond to the nearest faces of the analyzed one. The metric used is simply the Euclidean distance in the parameter space. For example in Figure 7, the vector will identify the face number as being the most similar to the analyzed one. The nearest models will correspond in the initial database to specific identities, which are the most similar to the identity of the unknown analyzed face.
4.2.3. Adapted Model
From this set of nearest identities, we generate an adapted database containing the corresponding faces in different expressions and poses. is a subset of the general database (Figure 5). Figure 8 illustrates such an adapted database when .
From , we generate the adapted model . When , 2, or 3, it is possible to evaluate beforehand the adapted model, depending on the number of different faces in the general database. For this database can contain up to one hundred faces, since the total number of combinations is around five thousands, and GB will then be sufficient to store the five thousand models. If then comparatively small general database will be used, that is, 33 different faces if only GB memory is available in the system.
When we need to align an unknown face in a static image, we then simply align the face with the initial model and apply the pre-computed model, which corresponds to the nearest faces. If a video stream related to one person needs to be analyzed, we use the first second of the stream in order to perform a more robust selection of the adapted model. On the first images, we align the face with the initial model . We evaluate the error (7) on each image. This error is remarkably stable, because of the use we make of the oriented map; it is then possible to compare it to a threshold, in order to decide if the model has converged. We then evaluate, from the correctly aligned faces, the nearest identities which must be taken into account in the general database, in order to construct the adapted model. This model is then used on the following images in the video stream, in order to align the face.
We will specify hereafter the parameters values and metric to evaluate the performances of our two contributions (OM-AAM and Adapted AAM). This section will end with a discussion on the different results.
5.1. Experiments Setup
We use the same metric as in , in order to evaluate the error,
where is the error made on one of the points representing the eyes, nose, and mouth centers; is the distance between the eyes. In the context of the robustness analysis to illumination, identity, and pose, those four points are sufficient to illustrate the performances of our proposals. The precision of the ground truth is roughly of the distance between the eyes of the annotated faces; beyond , we consider that the alignment is not correct. We will then evaluate the error in the range .
A texture of 1756 pixels is used, in association with a 68-key points shape model and we keep of the deformation, in order to reduce the appearance vector dimension. With regard to the oriented map, no specific parameterization is necessary: the orientation number () is quantified on height bits and is not related to the type of the testing base images.
5.2. OM-AAM Performances
Let us remember that our objective is to make the AAM robust to illumination variations without any increase in the processing time. The DM-AAM of  complies with our constraints. We then propose to illustrate the OM-AAM performances, in comparison to those of the DM-AAM and classical AAM. Those comparisons will be made in a generalization context: the faces used to construct the model (18 persons from the M2VTS database ) and the ones used for the tests come from distinct databases.
Most of the time, a process which increases the robustness of an algorithm in a specific case decreases its performances in standard cases . For that reason, we will test our suggestions on a database, which is dedicated to illumination problems (CMU-PIE: 1386 images of 66 faces under 21 different illuminations ) and on an other one representing different faces with several expressions taken in different backgrounds (BIOID: 1521 images ) under variable light exposition (see Figure 9). This latest database is more difficult to process, since the background can be different and the faces present various positions, expressions and aspects. People can have glasses, moustaches, or beard.
Figure 10 represents the percentage of the images, which have been aligned with the error (15). For example the point (0.15,0.8) on CMU results means that for of the test images, the centers of the mouth, eyes, and nose were detected with a precision less or equal to of the distance between the eyes of the analyzed face. The DM- AAMs are more powerful than the classical ones when used with normalized faces with variable illuminations (CMU-PIE database), but are useless in standard situations (BioId database). The DM-AAM uses a distance map, which is extracted from the image contours points. The threshold used to detect the contours point is crucially important, and is based on the assumption that all testing base images share the same dynamic. This is not the case of the BioId database, in which the image contrasts present a great variation. Conversely, OM-AAMs do not use any threshold, since we do not extract any edge information but the gradient information on each pixel of the image.
A reference point used in the state of the art technology is often the point of abscissa . On the CMU-PIE database, OM-AAMs are able to align of the faces with a precision less or equal to , when DM-AAM and classical ones are less efficient: their performances are, respectively, and . But when the faces are acquired in real situations, our proposition overcomes other methods: in the BIOID database, OM-AAM can align of the faces with a precision less or equal to , which represents a and performance gain, with regard to classical AAM and DM performances, respectively.
5.3. Adapted AAM Performances
We propose to test the adapted AAM on the static images of the general database (Figure 5). A test sequence is then made, with one unknown person presenting four expressions under five different poses; the learning base associated to this testing base is made of all the other persons. A cross-validation of type Leave-one-out is used. All faces are tested separately, using all the other ones for the learning base. All the faces of the database have been tested, representing at the end a set of 580 images with a big variety of poses, expressions, and identity. The initial database used to generate the initial model is the same as the one presented in Figure 6, apart from the fact that the testing face has been removed. It contains then 28 different faces. This model is applied on every single 20 images of the unknown face, in order to evaluate the nearest faces. Then the adapted model is finally applied on those 20 images in order to align them (detect the gravity center of the eyes, nose, and mouth). In order to analyze separately the benefits of the proposed algorithm, we use only classical normalized textures instead of oriented ones.
To be able to find the optimal parameter , we have tested our algorithm for different values within the range . Figure 11 shows the percentage of the face aligned with a precision less or equal to of the distance between the eyes, versus : the number of nearest faces. As we can see, in the range , the alignment performances are relatively stable. They collapse after ; the adapted model is based on fifteen faces in five poses and four different expressions. The parameter space is breaking up leading to a multi-manifold, and optimization becomes more difficult to conduct (cf. Section 3.3).
We compare the performances of our system when (Adapted AAM) to those of three others different AAM. The first one (AAM 28) gets identity as the only variability and is made of the 28 faces (the twenty-ninth being tested) in frontal view and neutral expression. The second one (AAM 560) is full of rich variability, since it is based on 560 images representing 28 faces, representing themselves four expressions under five different poses. Lastly the third one (AAM GM)  (see Section 2) uses Gaussian mixtures to specify the regions of plausible solutions in the parameter space (see Figure 13). It is interesting to compare our proposition to this method since it is dedicated to multi-manifold spaces. We cannot implement it on a restricted database like the one of "AAM 28" which represents only one cluster of frontal faces. Four Gaussians were used to catch the density on the 560 images of the rich database of "AAM 560" model. We use the three first components of the appearance vector as it was indicated by the authors since the density in the other dimensions is uniform.
5.4. Adapted AAM Performances Discussion
The algorithmic complexity of "Adapted AAM" and "AAM 28" is almost the same, since their appearance vector dimension is similar (around 25). Conversely, "AAM 560" and "AAM GM" are much more complex (appearance vector dimension around 250) and exclude a real-time implementation. As it was said in Section 3.2 the warping takes of total processing time for real-time implementation when dimension of parameter vector is less than 30 and small textures are used like the ones we implement in this paper. To be precise, the ten iterations used to align a face takes 9.3 ms on a P4-2GHz. Usually for real implementation, we test the AAM on three different scales and nine positions around the detected center of the face, so we need 251 ms to align the face. The results presented here use those different scales and positions. After one second we switch to tracking mode: only five positions are tested around the center of the face so the algorithm works at 21 Hz. If the dimension of the appearance vector (see (8)) is multiplies by ten, then the number of operations is rougthly multiplied by ten too, with the warping time being not affected by this dimension growth. Even in tracking mode, this increase will then lead to only a 2 Hz framerates for "AAM 560" or "AAM GM" which is not sufficient for real time applications.
Figure 12 shows the superiority of the "Adapted AAM" over the three other models. The performances of the "AAM 560" are less good than those of the "Adapted AAM." It is consistent with the fact that the database used to build the "AAM 560" is much more rich in variability: the parameter space of this latest model is split into multi-manifolds. The "AAM GM" is able to identify these manifolds but is still slightly less good as "AAM 560," that will be discussed hereafter.
If we look at the reference error (), then our proposition is ten times more rapid than the "AAM 560" because of the dimension of the appearence vector, and clearly more effective (performances improvment of ) than the same heavy "AAM 560" model. If we compare now the "Adapted AAM" to the other light model (AAM 28), the "Adapted AAM" has the same complexity and is more effective for of the images of the testing base. As a conclusion, our model is more rapid and effective than other models, because it has focused on a relevant database, which is related to the testing face.
To understand why the results of "AAM GM" are less good then the ones of "AAM 560" it is necessary to look at the trajectory used during the AAM convergence process. Figure 13 shows the four Gaussians which were found by the Expected Minimization algorithm to specify the density in the first three dimensions and the two trajectories of the solutions found by the two AAM during the convergence. Both of them are initialized in the middle of the space and of course have the same path in the beginning. After few iterations the "AAM GM" finds a solution in a region specified as empty and performs a gradient descent to go back in the best direction in a plausible solution region. For "AAM 560" part, it continues to reach the good cluster and nearby it tries to find the best solution. As illustrated by Figure 14 (a zoom on Figure 13), each time the classical process of AAM proposes a nonplausible solution to the "AAM GM," it tries to go back, for that reason the trajectory of the "AAM GM" is disturbed compare to the "AAM 560" smooth trajectory. In fact  contribution was very interesting but illustrated only on one image as performance evaluation: two very different shapes of one object were to be fined. Maybe if the initialization point is in a region specified as non plausible, then after one iteration only, the gradient descent on the density charaterized by the Gaussians leads to a very fast convergence in the good cluster region and then the classical optimization is able to find a nice solution.
The trajectories end of Figures 13 and 14 lead to the images (b) and (c) of Figure 15. The associated models are based on the same 560 learning examples with very rich variability, so they are both able to catch the orientation of the face. It is interesting to note that the shape found by the "AAM GM" (Figure 15(c)) is more natural than the one of Figure 15(b) which has a hard discontinuity in the chin region, but this shape is able to retrieve the good orientation when the "AAM GM" reproduces a frontal face which is logical since the test made during the optimisation process stacked the trajectories in the central cluster. In fact most of the time it is because the markers associated to the extremity of the nose are not well positioned for "AAM GM" (which influence the value of the error see Section 5.1) that this method is slightly less good than "AAM 560." Of course the "AAM 28" based on frontal faces as learning base is only able to retrieve frontal faces as illustrated in Figure 15(d). Adapted AAM (Figure 15(a)) is the only method capable to produce a shape without any discontinuity, in the good orientation and a well placed nose.
6. Conclusion and Perspectives
Active Appearance Models are very efficient to align known faces in constraints conditions (face pose and illumination). In order to make them robust to illumination variations, we have proposed a new AAM texture type and a new normalization during the optimization step. In order to make them robust to unknown faces moving in unknown poses in different expressions, we have suggested an adapted model. This adaptation is made by choosing, in a set of pre-computed models, the best suited model to the unknown face. Tests made on public and private databases have shown the interest of our propositions; it is now possible to align unknown faces in nonconstraint situations, with a precision, which is sufficient enough for most applications requiring an alignment process (face recognition, face gesture analysis, cloning). Unlike  (cf. Section 2), where a specific model is made out of the first image of a video stream, we seek for the model which is best suited to the unknown face. This difference is significant; an imperfect initial alignment has no definitive repercussions. Our system is then more robust in view of the errors made by the initial generic model. At last, it is to be noted that the Adapted-AAM with oriented texture offers the same computational complexity as the classical AAM; they can be implemented in real time.
For emotion analysis and lip-reading, it is necessary to have a very precise alignment in order to be able to track the face dynamic. Precisely, the alignment performances must be evaluated on the localization of several points around the eyes, eyebrows, and mouth and not only on their gravity centers. We are now working on an adapted and hierarchical AAM, which use for each face characteristics (eyes and mouth essentially), the most relevant adapted model.
Garcia C, Delakis M: Convolutional face finder: a neural architecture for fast and robust face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 2004,26(11):1408-1423. 10.1109/TPAMI.2004.97
Viola P, Jones MJ: Robust real-time face detection. International Journal of Computer Vision 2004,57(2):137-154.
Cootes TF, Edwards GJ, Taylor CJ: Active appearance models. Proceedings of the European Conference on Computer Vision (ECCV '98), 1998
Beichel R, Bischof H, Leberl F, Sonka M: Robust active appearance models and their application to medical image analysis. IEEE Transactions on Medical Imaging 2005,24(9):1151-1169.
Cootes TF, Taylor CJ: On representing edge structure for model matching. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '01), 2001 1: 1114-1119.
Scott IM, Cootes TF, Taylor CJ: Improving appearance model matching using local image structure. Proceedings of the International Conference on Information Processing in Medical Imaging (IPMI '03), 2003 258-269.
Stegmann MB, Larsen R: Multi-band modelling of appearance. Proceedings of the Workshop on Generative Model-Based Vision (GMBV '02), 2002
Hu C, Feris R, Turk M: Active wavelet networks for face alignment. Proceedings of the British Machine Vision Conference (BMVC '03), 2003
Giri D, Rosenwald M, Villeneuve B, Le Gallou S, Séguier R: Scale normalization for the distance maps AAM. Proceedings of the 9th International Conference on Control, Automation, Robotics and Vision (ICARCV '06), 2006, Singapore
Langs G, Peloschek P, Donner R, Reiter M, Bischof H: Active feature models. Proceedings of the International Conference on Pattern Recognition (ICPR '06), 2006 1: 417-420.
Darkner S, Larsen R, Stegmann MB, Ersboll BK: Wedgelet enhanced appearance models. Proceedings of the Workshop on Generative Model Based Vision (GMBV '04), July 2004, Washington, DC, USA
Zhu J, Liu B, Schwartz SC: General illumination correction and its application to face normalization. Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP '03), 2003 3: 133-136.
Huang Y, Lin S, Li SZ, Lu H, Shum H-Y: Face alignment under variable illumination. Proceedings of IEEE International Conference on Automatic Face and Gesture Recognition (FGR '04), 2004 85-90.
Zhao WY, Chellappa R: Illumination-insensitive face recognition using symmetric shape-from-shading. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '00), 2000 1: 286-293.
Georghiades AS, Belhumeur PN, Kriegman DJ: From few to many: generative models for recognition under variable pose and illumination. Proceedings of the International Conference on Automatic Face and Gesture Recognition (FGR '00), 2000
Lee K-C, Ho J, Kriegman DJ: Acquiring linear subspaces for face recognition under variable lighting. IEEE Transactions on Pattern Analysis and Machine Intelligence 2005,27(5):684-698.
Gross R, Matthews I, Baker S: Fisher light-fields for face recognition across pose and illumination. Proceedings of the German Symposium on Pattern Recognition, 2002
Christoudias CM, Morency L-P, Darrell T: Light field appearance manifolds. Proceedings of the European Conference on Computer Vision, 2004 481-493.
Kittipanya-ngam P, Cootes TF: The effect of texture representations on AAM performance. Proceedings of the International Conference on Pattern Recognition (ICPR '06), 2006 2: 328-331.
Romdhani S, Gong S, Psarrou A: A multi-view nonlinear active shape model using kernel PCA. Proceedings of the British Machine Vision Conference (BMVC '99), 1999 483-492.
Blanz V, Vetter T: A morphable model for the synthesis of 3D faces. In Proceedings of the International Conference on Computer Graphics and Interactive Techniques (SIGGRAPH '99), 1999. Addison-Wesley, Reading, Mass, USA; 187-194.
Li Y, Gong S, Liddell H: Modelling faces dynamically across views and over time. Proceedings of IEEE International Conference on Computer Vision, 2001 1: 554-559.
Xiao J, Baker S, Matthews I, Kanade T: Real-time combined 2D+3D active appearance models. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '04), 2004 2: 535-542.
Sattar A, Aidarous Y, Le Gallou S, Séguier R: Face alignment by 2.5D active appearance model optimized by simplex. Proceedings of the International Conference on Computer Vision Systems (ICVS '07), 2007
Dornaika F, Ahlberg J: Face model adaptation for tracking and active appearance model trainin. Proceedings of the British Machine Vision Conference (BMVC '03), 2003
Cootes TF, Walker KN, Taylor CJ: View-based active appearance models. Proceedings of the International Conference on Automatic Face and Gesture Recognition (FGR '00), 2000 227-232.
Cootes TF, Wheeler GV, Walker KN, Taylor CJ: Coupled-view active appearance models. Proceedings of the British Machine Vision Conference (BMVC '00), 2000 1: 52-61.
Li SZ, Shuicheng Y, Zhang HJ, Cheng QS: Multi-view face alignment using direct appearance models. Proceedings of the International Conference on Automatic Face and Gesture Recognition (FGR '02), 2002 324-329.
Hu C, Feris R, Turk M: Real-time wiew-based face alignment using active wavelets networks. Proceedings of the International Workshop on Analysis and Modeling of Faces and Gestures (AMFG '03), 2003
Feng X, Lv B, Li Z: Automatic facial expression recognition using both local and global information. Proceedings of the Chinese Control Conference, 2006 1878-1881.
Oost CR, Lelieveldt BPF, Üzümcü M, Lamb H, Reiber JHC, Sonka M: Multi-view active appearance models: application to X-ray LV angiography and cardiac MRI. Proceedings of the International Conference on Information Processing in Medical Imaging (IPMI '03), 2003 234-245.
Chang Y, Hu C, Turk M: Probabilistic expression analysis on manifolds. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '04), 2004 2: 520-527.
Hu C, Chang Y, Feris R, Turk M: Manifold based analysis of facial expression. Proceedings of the Conference on Computer Vision and Pattern Recognition Workshop (CVPRW '04), 2004 5: 81-87.
Christoudias CM, Darrell T: On modelling nonlinear shape-and-texture appearance manifolds. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '05), 2005 2: 1067-1074.
Cootes TF, Taylor CJ: A mixture model for representing shape variation. Image and Vision Computing 1999,17(8):567-573. 10.1016/S0262-8856(98)00175-9
Zalewski L, Gong S: 2D statistical models of facial expressions for realistic 3D avatar animation. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '05), 2005 2: 217-222.
Xu Z, Chen H, Zhu S-C: A high resolution grammatical model for face representation and sketching. Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR '05), 2005 2: 470-477.
Langs G, Peloschek P, Donner R, Bischof H: A clique of active appearance models by minimum description length. Proceedings of the British Machine Vision Conference (BMVC '05), 2005 859-868.
Tong Y, Wang Y, Zhu Z, Ji Q: Facial feature tracking using a multi-state hierarchical shape model under varying face pose and facial expression. Proceedings of the International Conference on Pattern Recognition (ICPR '06), 2006 1: 283-286.
Canzler U, Wegener B: Person-adaptive facial feature analysis. Proceedings of the International Conference on Electrical Engineering, 2004
Aidarous Y, Le Gallou S, Sattar A, Séguier R: Face alignment using active appearance model optimized by simplex. Proceedings of the 2nd International Conference on Computer Vision Theory and Applications (VISAPP '07), 2007, Barcelona, Spain 231-236.
Stegmann MB: Active appearance models: theory, extensions and cases, M.S. thesis. Informatics and Mathematical Modelling, Technical University of Denmark, DTU, Denmark; 2000.
Du B, Shan S, Qing L, Gao W: Empirical comparisons of several preprocessing methods for illumination insensitive face recognition. Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP '05), 2005 2: 981-984.
Zuiderveld K: Contrast Limited Adaptive Histogram Equalization. In Graphics Gems IV. Academic Press, Boston, Mass, USA; 1994.
Cristinacce D, Cootes TF: A comparison of two real-time face detection methods. Proceedings of the International Workshop on Performance Evaluation of Tracking and Surveillance, 2003
Froba B, Kublbeck C: Robust face detection at video frame rate based on edge orientation features. Proceedings of the 5th International Conference on Automatic Face and Gesture Recognition (FGR '02), 2002
Belaroussi R, Prevost L, Milgram M: Classifier combination for face localization in color images. Proceedings of the International Conference on Image Analysis and Processing (ICIAP '05), 2005 1043-1050.
Cristinacce D, Cootes TF: Feature detection and tracking with constrained local models. Proceedings of the British Machine Vision Conference (BMVC '06), 2006
Pigeon : M2VTS Project. M2VTS, 1996, http://www.tele.ucl.ac.be/PROJECTS/M2VTS/m2fdb.html
Sim T, Baker S, Bsat M: The CMU pose, illumination, and expression (pie) database. Proceedings of the International Conference on Automatic Face and Gesture Recognition (FGR '02), 2002
HumanScan AG BioID Technology Research : The bioid face database. 2001.http://www.bioid.com/downloads/facedb/