T. F. Cootes, C. J. Taylor, and A. Lanitis, "Active shape models: Evaluation of a multi-resolution method for improving image search," in Proc. Br. Machine Vision Conf., vol. 1, 1994, pp. 327--336.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... applicability for real life imagery (specifically, boundary detection in real time cross sectional echocardiography) Besides our current work on applying the technique to real medical data (Figure 7, Figure 8) we are also considering a multi resolution extension (similar to the one presented in [7]) and a time scaling and time translation feature. ....

Cootes T., Taylor C., Lanitis A., Active Shape Models: Evaluation of a MultiResolution Method for Improving Image Search. Proceedings of the British Machine Vision Conference, 1994, 327-336.

.... of a possible implementation to obtain the pyramid is to first smooth the image at the lower level with a 55 Gaussian mask (which is linearly decomposed into two 1 5 8 5 1 convolutions) see Figure 6) and then sub sampling every other pixel to obtain the image at one higher level (Cootes et al. [10]) 1 2 3 4 5 1 2 3 4 5 0 0.05 0.1 0.15 0.2 0.125 0.625 1 0.625 0.125 0.625 3.125 5 3.125 0.625 1 5 8 5 1 0.625 3.125 5 3.125 0.625 0.125 0.625 1 0.625 0.125 Figure 6. Gaussian mask The search will now be performed by first searching at the top level of the pyramid and then initiating the search ....

....the pyramid and n landmarks in each image. The mean is obtained, as before, by averaging the normalized profiles for a certain landmark along the N images of the training set. A criterion should be decided upon in order to change the level of search within the pyramid. One example (Cootes et al. [10]) is to move to a lower level when a certain percentage of the landmarks did not change considerably, for example when 95 of the landmarks move only within the central 50 of the search profile. A maximum number of iterations can also be devised to prevent from getting stuck at a higher level ....

Cootes T, Taylor C, Lanitis A. Active Shape Models: Evaluation of a MultiResolution Method for Improving Image Search. Proceedings of the British Machine Vision Conference, 1994, pp.327-336.

....140 160 180 200 50 100 150 200 250 300 50 100 150 200 250 300 350 400 100 200 300 400 500 600 Level: 1 Level: 0 Figure 1. Pyramid of Images In order to obtain the pyramid we first smooth the image at the lower level and then sub sample every second pixel to obtain the image at the higher level [3]. We start by searching at the top level of the pyramid and then continue at a lower level using the search output of the previous level. The procedure is repeated until the lowest level of the pyramid (the original image) is reached. In order to carry out this multi resolution search, we must use ....

....need a criterion for determining when to change the level of search within the pyramid. One possibility is to move to a lower level when a certain percentage of the land marks do not change considerably, for example when 95 of the landmarks move only within the central 50 of the search profile [3]. A maximum number of iterations can also be devised to avoid getting stuck at a higher level (See Figure 2 and 3) Progress in MR 10 15 20 5 10 15 20 25 30 35 Progress in HiR 100 200 300 400 100 200 300 400 500 600 Progress in MR 10 15 20 5 10 15 20 25 30 35 Progress in HiR 100 200 300 400 100 ....

T. Cootes, C. Taylor, A. Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search. Proceedings of the British Machine Vision Conference, 1994, pp.327336.

....require the search profile to be short for two reasons. First, to reduce the computations required, and second, if the search profile is long and the target point is close to the current position of the landmark then it will be more probable to move to a farther away point and miss the target. In [11] a multi resolution approach is suggested, where at first the search looks for far away points and makes large jumps, and as the search homes in on a target structure, it should restrict the search only to close points. In order to achieve such multi resolution search, we generate a pyramid of ....

....when segmenting ST shapes in new image sequences, is thought to give better segmentation results especially when the initial ST shape estimate is placed farther away from the target. Such improvement in performance has already been demonstrated in multi resolution ASM (refer to Section 4.3.2. 2 and [11]) and multi resolution deformable ST shapes can be formulated in a similar manner. 9.3. Deformable ST Shapes With Temporal Translations Extending the current work to include an extra degree of freedom accounting for temporal translation leads to the possibility of searching through longer image ....

Cootes T, Taylor C, Lanitis A. Active Shape Models: Evaluation of a MultiResolution Method for Improving Image Search. Proceedings of the British Machine Vision Conference, 1994, page(s): 327-336.

.... dimensions (Davatzikos and Prince, 1995) and for mapping the deep sulci in three dimensions using higher dimensional active ribbons (Le Goualher and Barillot, Other applications of similar models can be found in Sandor and Leahy (1997) Davatzikos and Prince (1996) Thompson and Toga (1996) Cootes et al. 1994), Cootes et al. 1994) and McInerney and Terzopoulos (1996) There have been several models proposed in the statistical literature for outlining shapes with parametrized boundary models, for example Grenander and Miller (1994) Clifford and Nicholls (1994) and Hurn and Rue (1997) Much of the ....

.... and Prince, 1995) and for mapping the deep sulci in three dimensions using higher dimensional active ribbons (Le Goualher and Barillot, Other applications of similar models can be found in Sandor and Leahy (1997) Davatzikos and Prince (1996) Thompson and Toga (1996) Cootes et al. 1994) Cootes et al. 1994), and McInerney and Terzopoulos (1996) There have been several models proposed in the statistical literature for outlining shapes with parametrized boundary models, for example Grenander and Miller (1994) Clifford and Nicholls (1994) and Hurn and Rue (1997) Much of the recent work has been on ....

Cootes, T. F., Taylor, C. J. and Lantis, A. (1994) Active shape models: evaluation of a multi-resolution method for improving image search. vol. 1, pp. 327--336. BMVA Press.

.... can be generated from a training set of examples [6] The PDM consists of a compact set of orthogonal shape parameters which can then be used for image fitting using the Active Shape Model (ASM) 7] Cootes et al. have extended the ASM using a multi resolution search strategy [8]. Baumberg and Hogg describe how the PDM approach can be extended to parametrised curves [9] By using a simple segmentation scheme training data can be automatically collected for model building [10] Principal Component Analysis (PCA) is used to generate orthogonal shape parameters and an ....

T F Cootes, C J Taylor, and A Lanitis. Active shape models: Evaluation of a multi-resolution method for improving image search. In British Machine Vision Conference, volume 1, pages 327--336, 1994.

.... have a reasonably high a priori probability density, assuming the training shapes were sampled from a Gaussian distribution about the mean shape (see Haslam et al. [23] In order to improve the speed and robustness of the ASM, a multi scale search mechanism can be used, described by Cootes et al. [24]. 2.2.3 Lowe re#nement Lowe describes an iterative scheme for #tting parametrised 3D models to images [22] The scheme is based on Newton s method and is stabilised using a priori constraints. Given a vector of nonlinear parameters p a sequence of estimates are calculated using p (i 1) p (i) ....

....contour to deform more easily in modes of variation that vary signi#cantly within the training set. The statistical framework can be used to automatically control the search scale for feature search on an individual frame (in a similar manner to the multi scale extension to the ASM of Cootes et al. [24]) as well as over successive frames 50 (allowing motion coherence to be exploited when #lock# has not been lost over the contour) A simple method is described to cope with known occlusion (e.g. when two tracked objects overlap in the image) improving the robustness of the system. A signi#cant ....

T F Cootes, C J Taylor, and A Lanitis. Active shape models: Evaluation of a multi-resolution method for improving image search. In British Machine Vision Conference, volume 1, pages 327#336, 1994.

....model more compactly and to reduce the possibility of illegal shape intensity parameter combinations. ffl The method currently requires an initial estimate of the lip position relatively close to the actual position. This requirement could be relaxed by the use of multi resolution image search [45]. This consists on a coarse to fine search strategy which also uses models of different resolution. The search is initialised at the coarsest resolution and progressively changes to smaller resolutions. Image search is therefore much faster and more robust. ffl To increase the robustness and ....

T. F. Cootes, C. J. Taylor, and A. Lanitis. Active shape models : Evaluation of a multi-resolution method for improving image search. In Proceedings of the British Machine Vision Conference, pages 327--336, 1994.

.... A similar approach was used by Cootes et al. to model the grey levels in regions around the model points of their Active Shape Models [4] They went on to demonstrate a multi resolution search technique in which grey level models were trained and used at each level of a gaussian image pyramid [5]. We have found that although the eigen feature models combined with the multi resolution search technique produce very promising results they occasionally fail unexpectedly. For instance, models trained to locate a surface mount resistor can occasionally (about 1 of the time) fit better to an ....

T.F. Cootes, C.J.Taylor, A.Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search, in Proc. British Machine Vision Conference, (Ed. E.Hancock) BMVA Press 1994, pp.327-338.

....as appearanceparameters. It is important to note that the coding we achieve is reversible a given face image can be reconstructed from its appearance parameters. When a new image is presented to our system, facial features are located automatically using Active Shape Model (ASM) search [5] based on the flexible shape model obtained during training. The resulting automatically located model points are transformed into shape model parameters. Grey level information at each model point is collected and transformed to local greylevel model parameters. Then the face is deformed to the ....

....set of appearance parameters; in the following sections we describe how we use these parameters to reconstruct and interpret face images. 5: Locating Facial Features The shape model and local grey level models de scribed above can be used to locate all the modelled fea tures simultaneously [4,5]. The mean shape model is placed in the image and is allowed to interact dynamically until it fits to the image evidence. Each iteration involves two main steps: calculating a new suggested Shape model fitted Shape extracted I Extracted profiles at model points Deform to mean face Fig 5: ....

[Article contains additional citation context not shown here]

TE Cootes, C.J. Taylor and A. Lanitis. Active Shape Mo- dels: Evaluation of a Multi-Resolution Method For Im- proving Image Search. Procs. of the 5th British Machine Vision Conference 1994.

....models can be used together to describe the overall appearance of each face in terms of its model parameters; collectively we refer to the model parameters as appear ance parameters. 3. 2 Identification Our flexible shape model can be used for locating facial characteristics automatically[ 6 ]. When a face image is presented to the system, an instance of the shape model is overlaid on the image. Forces are applied to model points, in an iterative scheme so as to move and deform the model, until it fits to the shape of the face presented. Based on the flexible shape model the resulting ....

....point were built using the method described in section 4. Most of Fig. 9: Extraction of a grey profile at a model point these models need 4 model parameters to explain 95 of the variation. 9 Locating Facial Characteristics The shape model described above can be used as an Active Shape Model[ 6 ] to fit to new faces in an iterative local optimization scheme. The face model is placed on the image and is allowed to interact dynamically until it fits to the shape of the face presented. During each iteration two main operations take place: the definition of a new suggested position for each ....

[Article contains additional citation context not shown here]

TF. Cootes, C.J.Taylor and A.Lanitis. Active Shape Mo- dels: Evaluation of a Multi-Resolution Method for Im- proving Image Search. To appear in the Procs. of the British Machine Vision Conference, 1994.

....of interest in each of a set of training images, aligning these training shapes into a common reference frame then performing a statistical analysis to obtain the mean shape and main modes of shape variation. They show how these Point Distribution Models (PDMs) can be used in image search [1,2] by creating Active Shape Models (ASMs) An ASM is analogous to a snake in that it refines its position and shape under the influence of image evidence, giving robust object location. Hill et al. [11] show how the PDM ASM approach can be extended to 3D when volume or range images are available, ....

T.F. Cootes, C.J.Taylor, A.Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search, in Proc. British Machine Vision Conference, (Ed. E.Hancock) BMVA Press 1994, pp.327-338.

....hand images displaying the same gesture will vary in form. We have approached these problems by modelling the ways in which the appearance of faces and hands can vary, using parametrised deformable models which take into account all the main sources of variability. A robust image search method [ 4 ] is used to fit the models to new face hand images and recover compact parametric descriptions. Given this compact coding, standard statistical pattern recognition techniques can be used to perform the task of interpretation. For our face interpretation experiments we have used the Manchester ....

....set of appearance parameters; in the following sections we describe how we use these parameters to code and interpret face images. 4 Locating Facial Features The shape model and local grey level models described above can be used to automatically locate all the modelled features simultaneously [ 3,4 ]. The mean shape model is placed in the image and is allowed to interact dynamically until it fits to the data. Each iteration involves two main steps: calculation of a new suggested position for each model point based on matching the local grey level models, followed by movement and deformation ....

[Article contains additional citation context not shown here]

T.F. Cootes, C.J. Taylor and A. Lanitis. "Active Shape Models: Evaluation of a Multi-Resolution Method For Improving Image Search". Procs. of the 5th British Machine Vision Con- ference 1994, vol 1, pp 327-336, ed. Edwin Hancock, BMVA Press, 1994.

.... of variation of model trained on above two face examples COMPARING PERFORMANCE OF MODELS FOR LOCATING IMAGE STRUCTURES The flexible models generated by the approach described above are identical in form to PDMs, and so can be used for image search in both the Active Shape Model (ASM) frame work [6 7] and the Genetic Algorithm approach described by Hill et al. [8] Here we com pare PDMs, FEMs and the combined models when used for local optimisation in an ASM. For an ASM we require a shape model describing the locations of landmark points, and a set of models of the grey levels in regions ....

....estimate the parameters of a model instance are iteratively refined so that each model point moves towards a nearby location suggested by its grey level model. This al lows fast and robust location of structures in unseen images. In the experiments below we use a multi resolution implementation [7]. Early iterations are performed on a low resolution image; as the search progresses the model automatically moves to higher and higher resolutions until convergence is reached. A set of 80 labelled images of faces (8 mugshots of 10 different people) was split into a training set and a test ....

TF. Cootes, C.J.Taylor, A.Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Approach to Improving Image Search. BMVC'94

....to fit to the image data more accurately. We present results for synthetic and real images and discuss how the methods can be used in an interactive bootstrap training scheme where problems with over constrained models are particularly important. 1 Introduction Active Shape Models (ASMs) [1,2] provide a useful means for locating objects in images. An ASM relies on having statistical models of the expected shape and grey level appearance of the object of interest, generated from training images. It represents an object as a set of model points, typically located on the object ....

.... , Other Found Points Figure 1: Because the shape model does Figure 2: A strong noise edge can attract not include a bending mode of vari one point, leading to a compromise solation, a least squares compromise posi ution for all the points. tion is found. A multi resolution search technique [2] can reduce the effects of spurious feature responses (by gradually reducing the length of the search profile) but does not help with an overconstrained shape model. In this paper we describe two complementary methods which allow the model points freedom to move away from their current model ....

T.F. Cootes , C.J.Taylor, A.Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search, in Proc. British Machine Vision Conference, (Ed. E.Hancock) BMVA Press 1994, pp.327-338.

....involves finding the nearest strong edge along a normal through the current point. Better results can be obtained by using a training set to build a statistical model to represent the image structure expected at each point. During search we use the model at each point to find the best nearby match [3]. In the second step we first update M to minimise the errors in the image plane, jX 0 M(x)j 2 . We then invert M to project X 0 into the model frame using x 0 = M 1 (X 0 ) If our model space is in the tangent space to some vector x t , then M(x) is defined to rotate, translate and ....

....converge. In practice, knowing the approximate position, scale and orientation is enough. The search can be attempted once for each mixture component, starting with the shape defined by the mean of each component. The method can be made more efficient and robust using multi resolution techniques [3], first searching on a coarse scale image and refining on finer and finer scale images. 8 5 Example: Brain Stem Model We have used the above approach to generate a model of the appearance of the brain stem in successive slices of an MR image of the head. Figure 9 shows a set of contours of a ....

T. F. Cootes, C. J. Taylor, and A. Lanitis. Active shape models : Evaluation of a multiresolution method for improving image search. In E. Hancock, editor, 5 th British Machine Vison Conference, pages 327--336, York, England, Sept. 1994. BMVA Press.

....as appearance parameters. It is important to note that the coding we achieve is reversible a given face image can be reconstructed from its appearance parameters. When a new image is presented to our system, facial features are located automatically using Active Shape Model (ASM) search [16,18] based on the flexible shape model obtained during training. The resulting automatically located model points are transformed into shape model parameters. Grey level information at each model point is collected and transformed to local grey level model parameters. Then the face is deformed to the ....

....compact and thus more specific. The first few significant modes of variation for both the linear PDM and MLP PDM are similar. We have carried out systematic experiments to compare the performance of linear PDMs and MLP PDMs in the context of their ability to locate facial features using ASM search [16, 18]. We have tested the fitting procedure by fitting the model to 40 face images. We performed two main experiments. For the first experiment the initial pose was chosen randomly within the following limits: rotation of 20 degrees, displacement from the correct position by 30 pixels, and ....

T.F. Cootes, C.J. Taylor and A. Lanitis. Active Shape Models: Evaluation of a MultiResolution Method For Improving Image Search. Procs. of the 5th British Machine Vision Conference 1994, 1, pp 327-336, ed. Edwin Hancock, BMVA Press, 1994.

....the overall appearance of each face in terms of its model parameters; collectively we refer to the model parameters as appearance parameters. Identification When a test image is presented to our system, facial characteristics are located automati# cally using an Active Shape Model (ASM) 8 ] based on the PDM shape model obtained during training. The resulting model points are transformed into shape model parameters. Grey level information at each model point is collected and transformed to local greylevel model parameters. Then the face is deformed to the mean face shape and the ....

....of faces can be performed using the flexible shape and grey level models described above. When a new face image is presented to the system the flexible shape model is used for locating the facial characteristics automatically, using a multi res# olution Active Shape Model (ASM) search [ 8 ] In our current system the user indicates the approximate position of the nose and the mean shape model is overlaid on the image; deformation and Euclidean transformations of the model are applied in an iterative scheme, until the model is fitted to the shape of the face presented. Examples of ....

T.F. Cootes, C.J. Taylor and A. Lanitis. Active Shape Models: Evaluation of a MultiResolution Method For Improving Image Search. Procs. of the 5th British Machine Vision Conference 1994, vol 1, pp 327-336, ed. Edwin Hancock, BMVA Press, 1994.

....information from the two aspects of the model, ASMs are constrained to deform only in ways in which the objects have been observed to deform in the training set global shape constraints are applied. A multi resolution version of this search process has been found to produce more accurate results [22]. ASMs have been applied successfully to image analysis problems in medicine [23 27] and industry [28] The details of both shape and gray level modelling have been described in [21] In this paper, we describe a study performed to test the feasibility of using an ASM to segment the femoral ....

....moved and deformed so as to better fit the image evidence. The whole process of searching around each model point for a better fit to its gray level model and updating the pose and shape of the model is repeated until no further improvement results. Multi Resolution Searching It has been shown [22], that the search technique described above can be improved by using a multi resolution approach, with models applied first, to a coarse, sub sampled version of the image, proceeding to higher resolution versions of the image. This allows the models to search over a wider area at the start, whilst ....

[Article contains additional citation context not shown here]

T. F. Cootes, C. Taylor, A. Lanitis, Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search. In 5 th British Machine Vision Conference (Ed. E. Hancock), 327--336, York, England (1994), BMVA Press.

....for translation, rotation, scale and shape parameters until convergence. During minimisation shape and grey level profile model parameters are constrained to lie within Sigma3s of the mean. For this paper we have also used a coarse to fine multiscale image search, initially used for ASM s in [7] but using a point wise iterative fit rather than a simplex minimisation over parameter space. For multiscale fitting each training image is successively Gaussian filtered and subsampled a number of times and a set of GLDM s are built, one for each scale. Each example image is likewise subsampled ....

T. F. Cootes, C. J. Taylor, and A. Lanitis. Active shape models: Evaluation of a multiresolution method for improving image search. In E. Hancock, editor, Proc. British Machine Vision Conference, pages 327--336, 1994.

....system we require a generic object search technique. We have used the Active Shape Model (ASM) technique described by Cootes et al. [3] They compare this approach with a number of other methods for image interpretation based on flexible or deformable models and identify the strength of the ASMs [4] as arising from their specificity to the particular object for which they are searching. Specificity gives an ASM robustness in the face of noise and clutter in the image but from our point of view the general applicability of the technique is even more important. This generality arises from ....

T. F. Cootes, C. J. Taylor, and A. Lantis, "Active shape models: Evaluation of a multi-resolution approach to improving image search." BMVC 94.

....Because the PDM imposes global shape constraints, only objects of similar shape to those observed in the training set will be located in images. A multiresolution approach, employing grey level models trained on a Gaussian pyramid of images, has been shown to improve speed and robustness [3]. We have employed this approach in our experiments. In addition, methods for shape deformation which allow more flexibility to move along a contour rather than perpendicular to it [6] have been used throughout. This has been shown to be particularly effective in aiding location of objects with ....

T. F. Cootes, C. Taylor, and A. Lanitis. Active Shape Models: Evaluation of a Multi-Resolution Method for Improving Image Search. In E. Hancock, editor, 5 th British Machine Vision Conference, pages 327--336, York, England, 1994. BMVA Press.

....an overall deformation of the shape in order to produce a better model fit. ASMs are constrained to deform only in ways in which the objects have been observed to deform in the training set global shape constraints are applied. A multi resolution version of this search process has been described [16] which can improve its accuracy. ASMs have been applied successfully to a range of image analysis problems in medicine [17 20] and industry [21] 5 Method 5.1 Building the Models For our experiments, 2D slices taken from 14 T 1 weighted 3D MR images of the knee were used to create an ASM. The ....

T. F. Cootes, C. Taylor, A. Lanitis, Active Shape Models: Evaluation of a MultiResolution Method for Improving Image Search. In 5 th British Machine Vision Conference (Ed. E. Hancock), 327--336, York, England (1994), BMVA Press.

No context found.

T. F. Cootes, C. J. Taylor, and A. Lanitis, "Active shape models: Evaluation of a multi-resolution method for improving image search," in Proc. Br. Machine Vision Conf., vol. 1, 1994, pp. 327--336.

No context found.

T. Cootes, C. Taylor, and A. Lanitis. Active shape models: Evaluation of a multi-resolution method for improving image search. In Proc. British Machine Vision Conf., pages 327--336, 1994.

No context found.

Cootes T., Taylor C., Lanitis A., Active Shape Models: Evaluation of a MultiResolution Method for Improving Image Search. Proceedings of the British Machine Vision Conference, 1994, 327-336.

No context found.

T. F. Cootes, C. J. Taylor, A. Lanitis: "Active Shape Models : Evaluation of a Multi-Resolution Method for Improving Image Search", Proc. British Machine Vision Conference, BMVA Press, 1994, pp. 327--336

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