T. F. Cootes, C. J. Taylor, A. Lanitis, D. H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. IEEE International Conference on Computer Vision,pages242--246, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... distract the tracker (figure 2) One method of overcoming this problem is to use statistical models of the intensity profiles capable of handling multi modal distributions to account for the intermittent presence of the teeth and tongue (e.g. by extending the appearance modelling of Cootes et al. [3]) An alternate solution is to use a Bayesian classification approach for feature detection as was done earlier using Fisher s linear discriminant. Modelling the distribution of colour pixel intensities was facilitated by the observation that there are three prominent components inside the mouth, ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. 4th Int. Conf. on Computer Vision, pp. 242--246, 1993.

....correspondence. Empirically,we have found that this links the two processes in a positive feedback loop. Iterating between the shape and texture steps causes the vectorized representation to converge after several iterations. Our vectorizer is similar to the active shape model of Cootes, et al. [17][16] 23] in that both iteratively fit a shape texture model to the input. But there are interesting differences in the modeling of both shape and texture. In our vectorizer there is no model for shape# it is measured in a data driven manner using optical flow. In active shape models, shape is ....

....First, pixelwise correspondence is computed between i std and i a , as indicated by the grey arrow. Shape a;std is a vector field that specifies a corresponding pixel in i a for each pixel in i std .Texture t a consists of the grey levels of i a mapped onto the standard shape. Cootes, et al. [17]) and face recognition (Craw and Cameron [18] 19] A relative shape measured with respect to a standard reference shape y std is simply the difference y a ; y std # whichwe denote using the shorthand notation y a;std . The relativeshapey a;std is the difference in shape between the individual ....

[Article contains additional citation context not shown here]

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proceedings of the International Conference on Computer Vision, pages 242--246, Berlin, May1993.

....parameters of this combination [2] Furthermore, it was not established how effective an objective function based on those particular quantities was in the first place. Since then, specific other features have been tried in various domains, e.g. matching intensity profiles at feature points [4], matching multiscale intensity and gradient at feature points [5] and centering within a region of classified pixels [6] But few, if any, comparative studies within a domain have been done. Let us consider what quantity should be optimized when seeking a shape in an image. To find the boundary ....

....the correct feature strength to reward. This must be found through statistical measurement, which is to say, training. This requires a ground truth training set, which is a collection of images, each with the desired contour correctly specified. Such training on image features has been done in [4], whose model learned a knD Gaussian of intensities perpendicular to a shape boundary at k feature points along line segments of n pixels. Within a similar framework, 5] used a Gaussian distribution of multiscale intensity and gradient at a few points around many organs simultaneously in a human ....

[Article contains additional citation context not shown here]

T. Cootes, A. Hill, C. Taylor, and J. Haslam, "Building and using flexible models incorporating grey-level information," in Proc. ICCV, pp. 242--246, May 1993.

....approach is demonstrated on synthetic as well as on real world image sequences showing moving hands with partial occlusions. 1 Introduction In an increasing number of application fields the objects to be modeled undergo deformations which have to be analysed and characterized. Deformable models [3, 4, 7] are mathematical models which incorporate knowledge about shapes and their variations. These models have been used with success in the analysis of still as well as dynamic images, to extract, track or characterize deformable objects. In this paper we introduce a modeling framework which relies ....

....set of representative shapes. kl analysis allows to approximate the global deformations of the original template by superimposing the main variations modes extracted from the shapes belonging to the learning set. Five to ten parameters are usually sufficient to obtain an accurate description. In [3], these global deformation parameters are adjusted to fit the model on edges extracted from the image. A deterministic relaxation scheme, which requires an initialization close to the optimal configuration, is used to find the deformation modes [3] In [4] Grenander et al. has obtained very ....

[Article contains additional citation context not shown here]

T.F. COOTES, C.J. TAYLOR, A. LANITIS, D.H. COOPER, and J. GRAHAM. -- Building and using flexible models incorporating grey-level information. -- In Proc. 4th Int. Conf. Comp. Vis., pages 242--246, Berlin, Germany, May 1993.

....to outliers, and use of quasi random points for efficiency. In Section 4 we successfully apply this algorithm to real video sequences of pedestrians as well as automated surveillance sequences. Section 5 concludes the paper. 2 Learning a linear human model The Point Distribution Model (PDM) [3] has proven to be a useful method for building a compact linear shape model from training examples of a class of shapes. The conventional PDM requires manual labelling of a set of points called the landmark points in each training image. These points are concatenated to form a shape vector and ....

....B spline with the control points placed at approximately uniformly spaced intervals along the contour is produced efficiently from each of these silhouettes. The control points of the B spline are then used as the landmark points in the PDM. We use techniques similar to that described in [1] and [3] with some improvements to build a linear human model. One improvement is in the parameterization of the B spline curve that is fitted to each extracted contour. Suppose that the set of points in a single human contour is Q k ,k = 0, m, and we want to approximate these points with a p th ....

T. F. Cootes, C. J. Taylor, A. Lanitis, D. H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. IEEE International Conference on Computer Vision,pages242--246, 1993.

....either sound or vision on its own. It is possible that audiovisual tracking could help not only with image transmission but also with improved noise rejection in the audio channel. 2 Active Contours Active Contours extend the application of pattern theory to object localisation in single images [11, 28, 14, 6], to image sequences. They also derive partly from work on dynamical systems of curves [8, 20] that interact with images. Possible object contours, chosen from a suitable space of curves, are hypothesised. This can be done deterministically as in a Kalman filter [12, 17] or, for cases of dense ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating greylevel information. In Proc. 4th Int. Conf. on Computer Vision, pages 242--246, 1993.

....using the techniques discussed above. First, a statistical shape model of a pedestrian was built using automatically segmented pedestrian contours from sequences obtained by a stationary camera (so that we can do background subtraction) We use well established computer vision techniques (see [22] and [23] to build a LPDM (Linear Point Distribution Model) We fit a NURB (Non Uniform Rational B spline) to each extracted contour using least squares curve approximation to points on the contour [21] The control points of the NURBs are then used as a shape vector and aligned using weighted ....

T.F.Cootes,C.J.Taylor,A.Lanitis,D.H.Cooper,andJ.Graham.Buildingand using flexible models incorporating grey-level information. Proc. IEEE International Conf. on Computer Vision, pp. 242-246, #993.

....eigenmodels are applied to model object deformations: in [19] natural shape recognition is based on eigenmodels, and, finally in [4] 5] 14] deformable model fitting is driven by projecting shape information in low dimensional spaces. Appearance and geometric information is integrated in [6]. We propose a gesture tracking and recognition system which is based on geometric and visual appearance. The key question is to combine several sources of variability (eigenspaces) These modal spaces are the core of the system. In the first section we define a general gesture model. In the ....

Cootes, T.F., Taylor, C.J., Lanitis, A., Cooper, D.H., Graham, J.: Building and Using Flexible Models Incorporating Grey-Level Information. In Proc. International Conference of Computer Vision. (1993).

....face images. In his study of illumination invariant recognition techniques, Hallinan ( Hallinan, 1995] describes deformable models of a similar general flavor. The work of Taylor and coworkers ( Cootes and Taylor, 1992] Cootes and Taylor, 1994] Cootes et al. 1992] Cootes et al. 1994] [Cootes et al. 1993]; Hill et al. 1992] Lanitis et al. 1995] on active shape models is probably the closest to ours. It is based on the idea of linear combinations of prototypes to model non rigid transformations within classes of objects. They use a very sparse set of corresponding points in their model (we ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In ICCV, pages 242--246, Berlin, May 1993.

....an object class from several examples with varying pose and lighting [22, 17] Betke [4] uses simulated annealing for fast 2D object recognition (tra#c signs) in noisy images by matching a new scene to a set of templates generated by transforming model images. In deformable template matching [28, 6], templates are constructed to model the non rigid features of the object, and at recognition time the templates are aligned to the image by minimizing an energy function for the individual features. These methods work well when the deformations are small they provide a detailed analysis of the ....

T. F. Cootes, C. J. Taylor et. al. "Building and Using Flexible Models Incorporating Gray Level Information", ICCV, 242-- 246, Berlin, 1993.

....recovered from observations in images with ground truth contours. For deformable models, work has been done on learning specialized shape and feature models, but has not been generalized. Among learned prior shape models, there have been a multidimensional Gaussian distribution of vertex positions [7]; Markov Random Fields of vertex displacements with respect to neighbors [12, 13] a Gaussian distribution of variations in vibration modes [16] and a Gaussian distribution in a Fourier harmonic representation [18] Among learned image features, there have been kn dimensional Gaussians of ....

.... distribution of variations in vibration modes [16] and a Gaussian distribution in a Fourier harmonic representation [18] Among learned image features, there have been kn dimensional Gaussians of intensities along line segments of n pixels perpendicular to a shape boundary at k feature points [7]; Gaussian distributions of multiscale intensity and gradient at a few points around many organs simultaneously [2] 3D models of shape incorporating the observed likelihood of nearby edges being spurious [3] histogram of pixel values from ground truth shape boundaries [9] and discriminant ....

[Article contains additional citation context not shown here]

T.F. Cootes, A. Hill, C.J. Taylor, and J. Haslam. Building and using flexible models incorporating grey-level information. In ICCV, pages 242--246, Berlin, May 1993. IEEE Computer Society.

....be parameterised by their control points. In practice this allows too many degrees of freedom for stable tracking and it is necessary to restrict the curve to a low dimensional parameter x, for example over an affine space [28, 5] or more generally allowing a linear space of non rigid motion [9]. Finally, probability densities p(x) can be defined over the class of curves [9] and also over their motions [27, 5] and this constitutes a powerful facility for tracking. Reasonable default functions can be chosen for those densities. However, it is obviously more satisfactory to measure the ....

.... of freedom for stable tracking and it is necessary to restrict the curve to a low dimensional parameter x, for example over an affine space [28, 5] or more generally allowing a linear space of non rigid motion [9] Finally, probability densities p(x) can be defined over the class of curves [9], and also over their motions [27, 5] and this constitutes a powerful facility for tracking. Reasonable default functions can be chosen for those densities. However, it is obviously more satisfactory to measure the actual densities or estimate them from data sequences (x 1 ; x 2 ; ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. 4th Int. Conf. on Computer Vision, 242--246, 1993.

....of interest, and segmenting the sets of points into coherent objects is challenging. An alternative is to use higher level information, whether by modelling objects with specific grey level templates [2] which may be allowed to deform [12] or with more abstract templates such as curved outlines [4, 5]. By including high level motion models [4, 1] these trackers can follow complex deformations in high dimensional spaces, but there tends to be a tradeoff between speed and robustness. Kalman filter based contour trackers which run in real time are very susceptible to distraction by clutter, and ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. 4th Int. Conf. on Computer Vision, 242--246, 1993.

....estimation of the displacement, however, is determined by a search only in the normal direction toward the strongest image edge [3] which therefore would not discriminate well between nearby edges and could fail to converge adequately. For this reason, the use of gray level models was advocated [2, 6]. Pentland and Sclaro# [9] uses linear deformations equivalent to the modes of vibration of the original shape. However, these modes are based on a generic elastic model that is not likely to be representative of the real variations which occur in a class of shapes. Cootes and Taylor [4] combine ....

T. F. Cootes, C. J. Taylor, A. Lanitis, D. H. Cooper and J. Graham, "Building and using flexible models incorporating grey-level information," Proc. Int. Fourth Conf. on Computer Vision, pp. 242-246, 1993.

.... = 1 M M X k=1 (X k Gamma X) X k Gamma X) T : Moreover, Principal Components Analysis (PCA) Rao, 1973] can be used to restrict the shape space S to explain most of the variance in the training set while keeping the dimension of S small, in the interests of computational efficiency [Cootes et al. 1993, Baumberg and Hogg, 1994, Lanitis et al. 1995, Beymer and Poggio, 1995, Baumberg and Hogg, 1995a, Vetter and Poggio, 1996] An example is given in figure 1. Figure 1. PCA for faces. A shape space of facial expressions is reduced here by PCA to the two dimensional space that best covers the ....

Cootes, T., Taylor, C., Lanitis, A., Cooper, D., and Graham, J. (1993). Building and using flexible models incorporating grey-level information. In Proc. 4th Int. Conf. Computer Vision, 242--246.

....(tongue touching gum ridge) place. We therefore would like to have a model which describes both, the shape of the inner and outer lips and the intensity around the mouth area. We use models based on point distribution models (PDM) also called active shape models (ASM) when used in image search [16, 15, 13, 14]. PDMs are flexible models which represent an object by a set of labeled points. The points describe the boundary or other significant parts of an object. The average shape and the principal modes of variation are captured from a labeled training set. The training examples need to be labeled in a ....

....around each model point and estimate their main modes of variation within a training set. 0 2 4 6 8 200 100 0 Grey Level Profile Points Figure 5: Grey level profile extraction. The grey level vectors are sampled perpendicular to the lip contour and centred at the model points. Following [15], we choose to sample one dimensional profiles g ij of length n p perpendicular to the contour and centered at point j for each training image i as shown in Fig. 5. But instead of calculating individual mean profiles and covariance matrices for each model point, we concatenate the profiles of 10 ....

T. F. Cootes, C. J. Taylor, A. Lanitis, D. H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proceedings of the International Conference on Computer Vision, pages 242--246, 1993.

.... example with image data and using differences between model and image to deform the shape (Active Shape Models [11] This paper describes how it is possible to model the grey levels expected at each point of the shape model, and how such grey level models can be used in Active Shape Model search [12,13]. We also present the results of systematic experiments to assess how well an ASM can locate the model points on objects in unseen images. Experiments have been performed using different types of grey level model and using different weighting schemes in the search algorithms. We compare the ....

....using each of the four types of profile model (grey derivative, normalised unnormalised) are examined below. 4 Active Shape Models We have previously described a method of fitting by local search given a starting approximation to the pose and shape parameters required to fit a model to an image [11,12,13]. By choosing a set of shape parameters b for a Point Distribution Model, we define the shape of a model object in an object centred co ordinate frame. We can create an instance, X, of the model in the image frame by defining the position, orientation and scale: X = M(s,O) x] Xc (14) where Xc ....

[Article contains additional citation context not shown here]

T.F. Cootes, C.J.Taylor, A.Lanitis,D.H.Cooper and J.Graham, Building and Using Flexible Models Incorporating Grey-Level Information, Proc. International Conference on Computer Vision, May 1993.

....give other deformations of the chamber bounda. The modelling technique has been applied successfully to a wide varieW of examples in cluding the ventricles of the brain in MR images [6] spinal vertebrae in X rays [24] the outline of the abdomen and the prostate in MR images (see below) hces [3], hands [1] chromosomes and industrial components. 4 Modelling Grey Level Appearance We wish to use our models for locating examples of objects in new images. For this pur pose, not only shape, but also grey level appearance is important. We account for this by examining the statistics of the ....

....in the slice it is much easier to find its boundary. 6.3 Other Examples We have applied the techniques described above to a number of different problems in both medical and industrial fields. For instance segmenting vertebrae in lateral X rays of the spine [24] and locating facial features [3]. In each case the same procedures were used to build the models and search images. 6.4 Active Shape Models Applied to 3 D and Time Sequence Images We have found that a simple 2 D PDM of the first and second ventricles of the brain can capture not only biological variation between individuals ....

[Article contains additional citation context not shown here]

Cootes, T F, Taylor, C J, Lanitis, A, Cooper, D H and Graham, J 'Building and Using Flexible Models Incorporating Grey-Level Information', Proc International Conference on Computer Vision, Berlin 1993, pp.242-246.

.... in 2D and 3D, to reduce operator time and human error on system start up and subsequently on batch transition, e) Finite Element Methods [9] for artificially creating parameter variability when, at start up, only few instances of a component are available, f) Grey Level Patch Tracking [6]. Methods for reducing the size of training sets are currently being studied at the Wolf son Unit and all aim at producing more specific models, coping with larger variability, reducing training time. At present, CHECK is implemented on UNIX based work stations, but a PC version is being ....

TF. Cootes, C.J.Taylor, A.Lanitis, D.H.Cooper and J.Graham, 1993, "Building and Using Flexible Models Incorporating Grey-Level Information", Proc. Fourth International Conference on Computer Vision. IEEE Computer Society Press, 242-246.

....of the deformation energy. 3 Point Distribution Models Point Distribution Models (PDM s) 4] are flexible models developed in an attempt to model shape variations of variable objects. So far they have been used for automatic image interpretation. Applications include locating both industrial [4,5] and biological [6] objects within images. However, this is the first time that PDM s have been used for classification. PDM s are generated from a set of training shapes belonging to the class of objects we wish to model. On each training shape a number of landmark points are placed at some key ....

....interact dynamically until it fits to the object in the image. At each model point a profile perpendicular to the boundary is extracted and a new preferred position for that point is estimated along the profile. Various ways of defining the new preferred position for each point have been developed [2,3,4,5]; the most common method is to choose the strongest edge along the profile since in many cases control points are located on boundaries. The key to the method is that model points do not move individually to the suggested new positions. The scale, translation, rotation and shape parameters are ....

[Article contains additional citation context not shown here]

TF. Cootes, C.J.Taylor, A.Lanitis, D.H.Cooper and J.Graham, "Building and Using Flexible Models Incorporating Grey Level Information", in Proceedings of the Fourth International Conference on Computer Vi- sion, 1993.

....Fig. 8. Extracting a grey profile at a model point It would be desirable to implement a multi resolution GA PDM in order to reduce the computational search time and possibly improve the performance. Using Grey Level Information We have investigated the use of local grey level profile models [ 3 ], for calculating the objective function. To train these models a number of training profiles are needed to model the grey level appearance at each model point. To obtain the training profiles, we overlay the training shapes on the corresponding training images and extract grey level profiles in a ....

....at each point a similarity measure (d) can be calculated. This procedure is illustrated in figure 9. By summing the best similarity measure (dmax) recorded for each model point, an overall estimate of the goodness of a model instance can be obtained. This procedure is described in detail elsewhere [ 3,11 ]. Running Multi Resolution GA PDM Search A possible problem with the approach described in the preceding section is the dependency of the search procedure on the length of extracted profiles (l in figure 9) If the lengths are short the image evidence used for assessing a solution may not be ....

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper and J. Graham. Building and Using Flexible Models Incorporating Grey Level Information. Procs of the 4th International Conference on Computer Vision, pp 242-246, IEEE Computer Society Press, 1993.

No context found.

T. F. Cootes, C. J. Taylor, A. Lanitis, D. H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In Proc. IEEE International Conference on Computer Vision,pages242--246, 1993.

No context found.

T. Cootes, C. Taylor, A. Lanitis, D. Cooper, and J. Graham. Building and using flexible models incorporating grey level information. In Proceedings of the International Conference on Computer Vision, pages 242--246, 1993.

No context found.

T. Cootes, A. Hill, C. Taylor, and J. Haslam, "Building and using flexible models incorporating grey-level information," in Proc. IEEE Int'l Conf. on Computer Vision, pp. 242--246, IEEE Computer Society, (Berlin), May 1993.

No context found.

T.F. Cootes, C.J. Taylor, A. Lanitis, D.H. Cooper, and J. Graham. Building and using flexible models incorporating grey-level information. In ICCV, pages 242--246, Berlin, May 1993.

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