L. Wiskott, "Labeled graphs and dynamic link matching for face recognition and scene analysis," Reihe Physik 53, 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....attempts have been made and the methods tend to Preprint submitted to Elsevier Preprint Fig. 1. Different salient facial feature points selected for establishing correspondence between face images. From left to right, correspondence is established between four [29] fourteen [22] thirty one [38] and thirty six [11] points respectively. rely on the computation of optical flow. They are inevitably computationally expensive and usually require human intervention in order to ensure accept able performance [2,36] Due to the difficulty and the likely prohibitive cost from computing dense ....

....a model and a novel image. The use of a mean shape avoids the need to compute many correspondence maps. However, matching still entails a relatively expensive search over parameters of variations in shape and texture. The match can be based on models of the local appearance of the landmarks [38] or along curves joining the landmarks [14,22,25] Crucially, linear active shape models are based on a number of implicit as sumptions: a) the shape of the object of interest can be defined by a relatively small set of explicit view models, b) the grey levels around a particular landmark are ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. PhD thesis, Ruhr-Universit/t Bochum, Germany, 1995.

....time processing is crucial. Fig. J. Successful matching of a model graph on the input image. Although not all nodes are positioned perfectly, the placing usually yields correct recognition. 3.1.2. Hand posture analysis When the moving hand stops its posture is analyzed by elastic graph matching [19, 36, 37]. Different hand postures are represented as attributed graphs, whose nodes carry local image information in the form of responses to Gabor based wavelets, while the edges (in the graph theoretical sense) contain geometrical information (figure 4) In previous work we have demonstrated that ....

Laurenz Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt am Main, 1996.

....complex components. The value f 1 ## 2 p is the spacing factor between kernels in the frequency domain with k max 1:7: The width of the Gaussian envelope function is given by s lkl with 2:5: For the comparison of jets two similarity functions with different properties have been proposed [11]: S abs a j a 0 j ############### S pha J 0 2 a j a 0 j cosf j 2 f 0 B B B B B 1 C C C C C A 3 S abs only uses the magnitudes of the complex filter responses, S pha also considers their phases. Both functions yield similarity values between ....

....changes when J 0 is moved across the image, S pha varies very rapidly because the phases of filter responses change significantly on a spatial scale corresponding to the wave vector k of strongly responding kernels. 2.1. Bunch graphs of hand postures The idea behind the bunch graph concept [10,11] is to express the variability in the jets of corresponding points in several images of an object (e.g. several left eyes of different persons) by labeling a node with a collection of jets rather than only a single jet. Here, we extend this idea to not only represent variations in the object ....

L. Wiskott, Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, PhD Thesis, vol. 53 of Reihe Physik, Verlag Harri Deutsch, Thun, Frankfurt a. Main, Germany, 1995.

....In contrast to [Leung et al. 95] the con guration of the local lter responses is not matched with a geometrical model. Instead, the global consistency of the pattern is veri ed by analyzing features at a coarse resolution. Detecting components has also been applied to face recognition. In [Wiskott 95] local features are computed on the nodes of an elastic grid. Separate templates for eyes, the nose and the mouth are matched in [Beymer 93, Brunelli Poggio 93] There are two interesting ideas behind part or component based detection of objects. First, some object classes can be described ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Ph.D. thesis, Ruhr-Universitat Bochum, Bochum, Germany, 1995. 24

....which the two jets refer. When comparing two jets during graph matching, the similarity between them is maximized with respect to d, leading to an accurate determination of jet position. This idea goes back to [Fleet and Jepson, 1990] Theimer and Mallot, 1994] We use it in the form developed by [Wiskott, 1995]. We are using both similarity functions, preferring the phase insensitive version, eq. 2.3) which varies smoothly with relative position, when first matching a graph, and using the phase sensitive version, eq. 2.4) when being interested in accurate positioning. 2.3 Elastic Graph Matching ....

....similarity function, eq. 2.4) which by utilizing the phase is very sensitive to small jet displacements. 2. 4 Elastic Bunch Graph Matching When attempting to find an as yet unknown face in an image and to define a graph to represent it, we make use of a data structure called a bunch graph [Wiskott et al. 1995]. It is similar to the graph as described above, but instead of attaching only a single jet to each node, we attach a whole bunch of jets, each derived from a different facial image. To form a bunch graph, a collection of facial images (the bunch graph gallery) is marked with node locations at ....

Wiskott, L. (1995): Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Reihe Physik vol. 53. Verlag Harri Deutsch, Thun, Frankfurt a. Main.

....The strong variation of the phases of the complex filter responses is used explicitly to compute the displacement with subpixel accuracy. Implemented as an iterative coarse to fine process, displacements up to half the wavelength of the kernel with the lowest frequency used can be computed (see [9] for details) Using jets with four different frequencies leads to a maximal displacement of 6 7 pixels. As already mentioned in the introduction, a much larger range would help only in the special case of a purely translational movement, in all other cases larger displacements are associated ....

L. Wiskott, Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, Verlag Harri Deutsch, Thun, Frankfurt a. Main, Reihe Physik, Vol. 53, 1995. This article was processed using the L a T E X macro package with LLNCS style

....vision research on the one hand and by attempts at modeling human vision on a neuronal level on the other hand. Invariant object recognition, e.g. becomes easy if a correspondence map of su#cient density and reliability can be constructed between objects and stored prototypes (Wurtz, 1995; Wiskott, 1996; Wiskott and von der Malsburg, 1996; Wurtz, 1997) The study (Konen and von der Malsburg, 1993) is another illustration of the power of such an algorithm. It describes a neural network based on Dynamic Link Matching (DLM) that learns to evaluate input patterns for the presence of one out of three ....

....one pointer to an input vector. Another possibility is Dynamic Link Matching (DLM) Konen et al. 1994) which also relaxes M1 such that there is in principle full connectivity between the planes, and the correct correspondences grow, while all others decline to small values. In (Wurtz, 1995; Wiskott, 1996; Wiskott and von der Malsburg, 1996) dynamic link matching is used to construct correspondence maps for face recognition. As self organization is a notoriously slow process the convergence speed is an important detail. Its importance may be lower in the typical application areas that are models ....

Wiskott, L. (1996). Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis.

....direction. These feature displacement estimates for each direction were combined into one displacement using a least squared error criterion. A simple measure of the saliency, of a feature point was de ned in terms of the amplitudes, a j , of the Gabor lter responses at that point [13]: s X j a 2 j (1) Displacements were only estimated for salient feature points i.e. those with larger than a prede ned threshold. An initial estimate of a feature s displacement was based upon the phase responses, j , of the Gabor wavelets at the lowest spatial frequency level used ....

....practice, only low frequency Gabor responses were needed because the phase information yielded sub pixel accuracy. The estimated feature displacement, d, was assigned a con dence measure proportional to the similarity S(J; J 0 ) between the jets before (J 0 ) and after (J) displacement, where [13]: S(J; J 0 ) P j a j a 0 j cos( j 0 j dk j ) q P j a 2 j P j a 02 j (2) Features on the face which typically had high saliency were selected for tracking. In the current implementation, feature positions were initialised by hand in the rst frame of each sequence. ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. PhD thesis, Ruhr-Universitat Bochum, July 1995.

....absolute value a j and phase OE j according to c j = a j e iOE j . The filters respond strongly to edge and bar like structures, whose dimensions cover those typically produced by the teacher s grasping fingers. The similarity between two Gabor jets can be computed with two similarity functions [13]. The first uses only the absolute values a j of the complex filter responses, thereby disregarding their phases: S abs (J; J 0 ) P j a j a 0 j q P j a 2 j P j a 0 j 2 : 3) The second similarity function takes the filter responses phases into account: S pha (J; J 0 ) ....

....of the displacement vector d and the wave vector k j of the kernel. Now the similarity is maximized with respect to d by expanding the cosine into a Taylor series, computing the derivative with respect to d, and solving the resulting system of linear equations. Details on this can be found in [13]: S disp (J; J 0 ; d) P j a j a 0 j i 1 Gamma 1=2 Gamma OE j Gamma OE 0 j Gamma dk j Delta 2 j q P j a 2 j P j a 0 j 2 : 7) This gives an estimate of d with subpixel accuracy. The method has already been successfully applied to tracking faces in image sequences ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, volume 53 of Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt a. Main, 1995. PhD thesis.

....image I. Here we give a informal description of this total similarity; a precise mathematical description can be found in [WFK, KPM] Firstly, we define a similarity between two jets or jet similarity. We use two different jet similarities, the magnitude similarity [LVB] and the phase similarity [Wis]. The phase similarity is more sensitive to spatial displacements and gives as additional information to the similarity value an estimate of the displacement which allows more precise matching of landmarks. Utilizing the similarities of jets we define a local similarity between a jet extracted at ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Verlag Harry Deutsch (Frankfurt am Main) 1995.

....these ambiguities is taking the spatial arrangement of points into account in addition to the feature similarities. In the following, I will briefly describe a class of dynamic link networks that can solve the correspondence problem for human faces (Wurtz, 1995; Wiskott and von der Malsburg, 1996; Wiskott, 1996; Wurtz and von der Malsburg, in preparation) Image and model (the actual and the memorized face) are both represented by a sheet of neurons, each of which carries a local feature vector (a hypercolumn of simple cells) with it. Both sheets are fully interconnected by dynamic links, i.e. each ....

....other ones. This system converges to a one one mapping between the layers that reflects the correct correspondences. Variations on this theme are described by von der Malsburg (1988) Lades et al. 1993) Konen and Vorbruggen (1993) Wurtz (1995) Wurtz and von der Malsburg (1996, in preparation) Wiskott (1996) and Wiskott and von der Malsburg (1996) Konen and Vorbruggen (1993) and Wiskott and von der Malsburg (1996) extend the above system to a complete object recognition system by means of competition between several object layers. Wurtz (1995) Neural Networks: What is Lacking 109 and Wurtz and ....

Wiskott, L. (1996), Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, Reihe Physik, Verlag Harri Deutsch, Thun, Frankfurt am Main.

....time processing is crucial. Fig. 4. Successful matching of a model graph on the input image. Although not all nodes are positioned perfectly, the placing usually yields correct recognition. 3.1.2. Hand posture analysis When the moving hand stops its posture is analyzed by elastic graph matching [19, 36, 37]. Di#erent hand postures are represented as attributed graphs, whose nodes carry local image information in the form of responses to Gabor based wavelets, while the edges (in the graph theoretical sense) contain geometrical information (figure 4) In previous work we have demonstrated that elastic ....

Laurenz Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt am Main, 1996.

....similarity zero for our images, which have eight bits of resolution in intensity. Here, reliability of an intensity value is the same at all locations, so we can set c(i) 1 everywhere. 3. 2 Motion To extract the flow field we use the method described by Theimer Mallot [3] and by Wiskott [9]. We filter images with a Gabor wavelet transform for motion detection, which is based on the idea that a motion in image space results in a phase shift of the corresponding filter responses (see [3] for details) The resulting motion vector measured at a point x is called d( x) The ....

....above, motion detection at the edges of objects is often not very accurate. It is possible to judge the motion detector s confidence by comparing the filter responses at the original position x with those at the estimated new position x 0 = x d( x) with an appropriate similarity function [9]. 4 Object Recognition as a Model Based Cue We use an object recognition system described by Wiskott von der Malsburg [10] which is able to deal with mutual occlusion to a large extent and shows very good performance even in complicated scenes (see Fig. 1E for a typical example) It is based ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. PhD thesis, Ruhr-Universitat Bochum, 1995.

....small displacements only, i.e. large overlap of the Gabor jets, large displacement vectors are treated as a first estimate only and the process is iterated. We used four iterations. In this way displacements up to half the wavelength of the kernel with the lowest frequency can be computed (see [16] for details) For each vertex of the graph of frame n the displacements are calculated for frame n 1. Then a graph is created with its vertices at the new corresponding positions in frame n 1, and the labels of the new vertices are extracted from the new positions. But although the ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, volume 53 of Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt am Main, 1995. Acknowledgement We like to thank Michael Neef for system administration, Uta Schwalm for o#ce administration, and Pervez Mirza for proof-reading.

....application in face recognition. Our approach to solve the facial expression recognition problem is to take full advantage of face recognition technologies. Specifically, we try to apply the Elastic Graph Matching method which has been successfully developed in our FLAVOR face recognition system[6, 9, 10] to the difficult task of facial expression recognition. Although face recognition and facial expression recognition are different tasks, they are quite closely related. Both tasks may share some important characteristics in terms of the representation of faces, the method to identify feature ....

....an expression is about 8 seconds. However, the judgement about expression is made on a single face image rather than a video sequence. 2. Elastic Graph Matching Since Elastic Graph Matching has been applied to our system, it is helpful to give a short review of it. For details, please refer to [9]. In Elastic Graph Matching method, the basic visual feature of the image is described in the form of a vector called jet . Each component of a jet is a filter response of an image point with regard to a Gabor wavelet of specific frequency and orientation. The Gabor wavelet consists of a ....

L. Wiskott. Labeled graphs and dynamic link matching for face recognition and scene analysis. Verlag Harri Deutsch, Thun, Frankfurt a. Main, Reihe Physik, 53, 1995.

....the positions of local maxima in the left and right attention map. For the purpose of this example procedure, we only compute the spatial position of the object giving the maximum response. When tracking detects a stopping of the hand, its posture is analyzed by elastic graph matching (EGM) [5, 10]. EGM is a neurally inspired framework for object recognition, where objects, in this case different hand postures, are represented as labeled graphs (Fig. 1 center) where the nodes carry local image information in the form of responses to Gabor based wavelets, while the edges contain geometrical ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, volume 53 of Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt am Main, 1995.

....of its nodes identify landmarks. To calculate graph similarities a similarity measure between two jets has to be defined. At different stages of the match procedure we use two different jet similarity measures, magnitude similarity S mag (J ; J 0 ) 11] and phase similarity S pha (J ; J 0 ) [15, 16]. The phase similarity is more sensitive to spatial displacements and gives, in addition to the similarity value, an estimate of the displacement d which allows more precise matching of landmarks. As a node is actually labelled with a bunch of jets, we define bunch similarities (S B mag (B; ....

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. Verlag Harri Deutsch, Reihe Physik, Vol. 53, 1995.

....successive frames of a sequence, disparities between pairs of stereo images, or small positional displacements in a matching procedure are basically all the same task. The algorithm presented here was adapted from a stereo algorithm (Theimer Mallot, 1994) and has been used for matching purposes (Wiskott, 1995) as well. Consider two jets J , J 0 taken from the same pixel position in two successive frames. The underlying object in the scene may have moved by a vector d. The phases of the jet coefficients J j then vary mainly according to their corresponding wave vectors k j , yielding phase differences ....

....j ) are weighted by the amplitudes a j , because strongly responding jet coefficients are supposed to provide more reliable phases. The denominator normalizes S OE to the range of [ Gamma1; 1] so that it is independent of the contrast of the images. S OE can be used as a similarity function (Wiskott, 1995) and provides a cue as to how reliable the displacement estimation is. The function S OE is maximized in its Taylor expansion S OE (J ; J 0 ) P j a j a 0 j [1 Gamma 0:5( DeltaOE j Gamma dk j ) 2 ] r P j a 2 j P j a 02 j : 6) Setting dx S OE = dy S OE = 0 and ....

Wiskott, L. (1995). Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, volume 53 of Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt a. Main, Germany. PhD thesis.

....a composition of graphs of identical structure. In this structure, each node is labeled with a set of jets. By this means, the matching can choose between alternative jets for each node to achieve higher precision and be more generally applicable. The matching is described in detail elsewhere [9]. In this paper I focus on analyzing the matching result, generating composite or phantom faces, and determining facial attributes. Supported by grants from the German Federal Ministry for Science and Technology (413 5839 01 IN 101 B9) and from the US Army Research Laboratory (01 93 K 0109) ....

....rapid phase shifts. d is determined by maximizing S OE in its Taylor expansion around d = 0, which is a constraint fit of the two dimensional d to the 40 phase differences OE j Gamma OE 0 j (a detailed description of the estimation of the displacement vector d can be found elsewhere [9]) The jets and the similarity function are robust against changes in lighting conditions in two respects. Firstly, since the kernels are DC free, the jets are invariant with respect to general offsets in the image grey values. Secondly, since the similarity function S OE is normalized, it is ....

[Article contains additional citation context not shown here]

L. Wiskott, Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, vol. 53 of Reihe Physik. Thun, Frankfurt am Main, Germany: Verlag Harri Deutsch, 1995. (PhD thesis).

....become graphs, in which nodes are labeled with jets and edges are labeled with distance vectors. A small set of manually controlled model graphs constitute a face bunch graph. It represents the face space and is used to generate graphs of new faces by elastic graph matching (Lades et al. 1993; Wiskott et al. 1995), the algorithmic counterpart of dynamic link matching. In this work I only focus on analyzing the matching result, generating composite or phantom faces, and determining facial attributes. 2 Face Representation We use graphs with an underlying two dimensional topography. The nodes are labeled ....

....Gabor kernel and d is an estimated displacement vector which compensates for the rapid phase shifts. d is determined by maximizing S OE in its Taylor expansion around d = 0, which is a constraint fit of the two dimensional d to the 40 phase differences OE j Gamma OE 0 j , see (Wiskott 1995) for a detailed description. This preprocessing was chosen for its technical properties and its biological relevance. Since the wavelets have a limited localization in space and frequency, they are robust to shift, scaling, and rotation. The jets and the similarity function are robust against ....

[Article contains additional citation context not shown here]

Wiskott, L. (1995). Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, volume 53 of Reihe Physik. Verlag Harri Deutsch, Thun, Frankfurt a. Main, Germany. PhD thesis.

No context found.

L. Wiskott, "Labeled graphs and dynamic link matching for face recognition and scene analysis," Reihe Physik 53, 1995.

No context found.

L. Wiskott, "Labeled graphs and dynamic link matching for face recognition and scene analysis," Reihe Physik, 53, 1995.

No context found.

L. Wiskott, Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis, vol. 53 of Reihe Physik, Verlag Harri Deutsch, Thun, Frankfurt am Main, Germany, 1995. (PhD thesis).

No context found.

L. Wiskott. LabeledGraphs and Dynamic Link Matching for FaceRecognition and Scene Analysis. Ph.D. thesis, Ruhr-Universitat Bochum, Bochum, Germany, 1995. 24

No context found.

L. Wiskott. Labeled Graphs and Dynamic Link Matching for Face Recognition and Scene Analysis. PhD thesis, Ruhr-Universitat Bochum, Bochum, Germany, 1995.

First 50 documents

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