## Determining the Similarity of Deformable Shapes (1995)

Venue: | Vision Research |

Citations: | 107 - 7 self |

### BibTeX

@ARTICLE{Basri95determiningthe,

author = {Ronen Basri and Luiz Costa and Davi Geiger and David Jacobs},

title = {Determining the Similarity of Deformable Shapes},

journal = {Vision Research},

year = {1995},

volume = {38},

pages = {135--143}

}

### Years of Citing Articles

### OpenURL

### Abstract

We study how to measure the degree of similarity between two image contours. We propose an approach for comparing contours that takes into account deformations in object shape, the articulation of parts, and variations in the shape and size of portions of objects. Our method uses dynamic programming to compute the minimum cost of bringing one shape into the other via local deformations. Using this as a starting point, we investigate the properties that such a cost function should have to model human performance and to perform usefully in a computer vision system. We suggest novel conditions on this cost function that help capture the part-based nature of objects without requiring any explicit decomposition of shapes into their parts. We then suggest several possible cost functions based on different physical models of contours, and describe experiments with these costs. 1 Introduction Detecting similarity is a key tool in interpretating images. In this paper we develop a measure of s...

### Citations

8786 |
Introduction to Algorithms
- Cormen, Leiserson, et al.
- 1992
(Show Context)
Citation Context ...ely on dynamic programming, they may in some cases be more efficient than the dynamic programming method we have described, and we find this to be true on our problem (see Cormen, Leiserson and Rivest=-=[15]-=- for more on this difference). 24 eq (3) real data 10.52 9.16 7.78 7.21 6.61 5.41 synthetic data 11.71 11.47 11.23 11.01 10.80 10.60 eq (4) real data 11.74 9.93 9.47 8.66 8.31 6.63 synthetic data 12.2... |

4053 |
Pattern Classification and Scene Analysis
- Duda, Hart
- 1973
(Show Context)
Citation Context ... 12, 13]. Another approach to shape comparison describes shapes with a list of properties and their relations, using pattern matching techniques to determine similarity (eg. see [27], and a review in =-=[10]). A third-=- class of methods attempts to assign "semantic interpretation" to shapes, such as function, and make similarity judgements based on semantics (eg. [37, 16, 29, 28]). Also relevant to our wor... |

3892 |
Stochastic Relaxation, Gibbs Distribution, and the Bayesian Restoration of Images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...lows for articulations is a function that identifies the position of extreme bending and assigns a constant cost to these bendings. A cost function that achieves that, in the spirit of Geman and Geman=-=[18]-=-, Mumford and Shah[38], and Blake and Zisserman[8], can be written (in a discretized form) as E bending = P S s=1 (k t t 0 \Gamma k s ) 2 (1 \Gamma l s ) + fll s ; where l s = 0; 1. l s = 1 occurs if ... |

3184 | Snakes: Active contours models
- Kass, Witkin, et al.
- 1987
(Show Context)
Citation Context ...arping or gradient descent for computing the amount of deformation required. There has been a great deal of work on this approach. For example, it has been applied to tracking non-rigid objects (eg., =-=[14, 19, 8, 24]-=-) where the amount of deformation between frames may be assumed to be relatively small. This approach has also been widely applied to handwriting recognition (eg., [30, 38, 33]) using measures of simi... |

1005 |
Features of similarity
- Tversky
- 1977
(Show Context)
Citation Context ... Our work uses the intuition that as L 2 decreases, the two figures on the right each become more similar to the figure on the left. be quite similar to both, violating the triangle inequality. Also, =-=[34]-=- argues that human perception does not obey the symmetry constraint. Given our formulation of C (Eq. 1), C is a metric if and only if its integrand, Fds, is too. This implies: 1. F (k 1 ; k 2 ; t 0 )s... |

750 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...s the position of extreme bending and assigns a constant cost to these bendings. A cost function that achieves that, in the spirit of Geman and Geman[18], Mumford and Shah[38], and Blake and Zisserman=-=[8]-=-, can be written (in a discretized form) as E bending = P S s=1 (k t t 0 \Gamma k s ) 2 (1 \Gamma l s ) + fll s ; where l s = 0; 1. l s = 1 occurs if the difference in matched curvature is high compar... |

595 |
Feature extraction from faces using deformable templates
- Yuille, Hallinan, et al.
- 1992
(Show Context)
Citation Context ...es of similarity that tend to be particularly crafted to that domain. Deformable templates that reflect the specific shape of known objects have also been applied to contour extraction and modelling (=-=[39, 31, 21]-=-). Of particular relevance to our work: [32] suggest such an approach to judging the similarity of objects; [1] also propose such a method of comparing shape, with particular attention to producing fa... |

503 | Comparing images using the Hausdorff Distance
- HUTTENLOCHER, Klanderman
- 1993
(Show Context)
Citation Context .... Pentland and Sclaroff[40] construct mass and stiffness matrices for the given shapes, and deform the shapes by aligning the principal modes of their mass and stiffness matrices. Huttenlocher et al.,=-=[26, 27]-=- compare two shapes by applying the rigid (or affine) transformation that minimizes the Hausdorf measure between them. To allow for occlusion only a certain fixed fraction of the contour points are us... |

470 |
Representation and recognition of the spatial organization of three dimensional shapes
- MiuT, Nishihara
- 1978
(Show Context)
Citation Context ...will be used to determine the class identity ("categorization") of objects and to cluster objects into classes. A good deal of work on categorization has emphasized the part structure of obj=-=ects (eg. [5, 6, 23, 7, 2, 25, 9, 20]-=-). We believe that part structure plays an important role in categorization, but we believe also that metric information has a significant influence on our perception of shapes. As an example, conside... |

395 |
Time warps, string edits, and macromolecules: The theory and practice of sequence comparison
- Sankoff, Kruskal
- 1983
(Show Context)
Citation Context ...ld note that elastic matching has been widely used outside of vision to compare 1-D objects, and especially to compare speech signals. A review of some of this work can be found in Sankoff and Kruskal=-=[43]-=-. Many of these methods simply propose an intuitively appealing cost function, which may be crafted for a particular application domain. However, some papers also attempt to determine general properti... |

358 |
Dynamic 3D Models with Local and Global Deformations : Deformable Superquadrics
- Metaxas
- 1991
(Show Context)
Citation Context ...es of similarity that tend to be particularly crafted to that domain. Deformable templates that reflect the specific shape of known objects have also been applied to contour extraction and modelling (=-=[39, 31, 21]-=-). Of particular relevance to our work: [32] suggest such an approach to judging the similarity of objects; [1] also propose such a method of comparing shape, with particular attention to producing fa... |

241 |
Boundary detection by minimizing functional
- Mumford, Shah
- 1985
(Show Context)
Citation Context ... is a function that identifies the position of extreme bending and assigns a constant cost to these bendings. A cost function that achieves that, in the spirit of Geman and Geman[18], Mumford and Shah=-=[38]-=-, and Blake and Zisserman[8], can be written (in a discretized form) as E bending = P S s=1 (k t t 0 \Gamma k s ) 2 (1 \Gamma l s ) + fll s ; where l s = 0; 1. l s = 1 occurs if the difference in matc... |

239 | Parts of recognition - Hoffman, Richards - 1984 |

223 |
Human Image Understanding: Recent Research and a theory
- Biederman
- 1985
(Show Context)
Citation Context ...will be used to determine the class identity ("categorization") of objects and to cluster objects into classes. A good deal of work on categorization has emphasized the part structure of obj=-=ects (eg. [5, 6, 23, 7, 2, 25, 9, 20]-=-). We believe that part structure plays an important role in categorization, but we believe also that metric information has a significant influence on our perception of shapes. As an example, conside... |

194 |
Symbolic reasoning among 3-D models and 2-D images
- Brooks
- 1981
(Show Context)
Citation Context ...will be used to determine the class identity ("categorization") of objects and to cluster objects into classes. A good deal of work on categorization has emphasized the part structure of obj=-=ects (eg. [5, 6, 23, 7, 2, 25, 9, 20]-=-). We believe that part structure plays an important role in categorization, but we believe also that metric information has a significant influence on our perception of shapes. As an example, conside... |

179 |
Visual perception by computer
- Binford
- 1971
(Show Context)
Citation Context |

177 |
Object matching using deformable templates
- Jain, Zhong, et al.
- 1996
(Show Context)
Citation Context ... by measuring the deformations needed to relate the 4 regions bounded by the shapes' contours, rather than the contours themselves (eg., Amit, Grenander and Piccioni[1] and Jain, Zhong, and Lakshmanan=-=[28]-=-). In Section 7 we briefly discuss some of the advantages and disadvantages of comparisons based on regions rather than the contours that bound them. Another approach is to develop a general model of ... |

171 |
Closed-form solutions for physically based shape modeling and recognition
- Pentland, Sclaro
- 1991
(Show Context)
Citation Context ...A third class of methods attempts to assign "semantic interpretation" to shapes, such as function, and make similarity judgements based on semantics (eg. [37, 16, 29, 28]). Also relevant to =-=our work, [26]-=- deform shapes by aligning the principal modes of their mass and stiffness matrices, [40] combine part and deformable template methods in a novel way quite different from ours, [18] compare shapes by ... |

169 |
Aligning pictorial descriptions: An approach to object recognition
- Ullman
- 1989
(Show Context)
Citation Context ...heir mass and stiffness matrices, [40] combine part and deformable template methods in a novel way quite different from ours, [18] compare shapes by minimizing the Hausdorff measure between them, and =-=[35]-=- proposes aligning shapes by applying piecewise affine transformations. See [3] for further discussion of this work. Our approach: We consider the applicability of continuous-deformation methods to th... |

165 |
Pattern classification and scene analysis, A Wiley Iinterscience Publication
- Duda, Hart
- 2000
(Show Context)
Citation Context ...son is to describe shapes with a list of properties (features). These properties may be global (e.g., "the object is polygonal") or local. When global properties are used (see reviews in Dud=-=a and Hart[16]-=- and Bolles and Cain[9]) objects are represented as points in feature space. This representation is obtained by arranging the list of properties of an object as the components of a vector associated w... |

162 | Training models of shape from sets of examples
- Cootes, Taylor, et al.
- 1992
(Show Context)
Citation Context ...oth contours that fit the gradients of the image as well as possible. Yuille, Cohen, and Hallinan[57] recognize parts of a face using deformable models that they have hand-crafted, while Cootes et al.=-=[14]-=-, Baumberg and Hogg [5], and Hinton, Williams and Revow[23] suggest learning deformable models of objects from examples. Finally, we should note that elastic matching has been widely used outside of v... |

157 |
Dynamic programming for detecting, tracking, and matching deformable contours
- Geiger, Gupta, et al.
- 1995
(Show Context)
Citation Context ...trings, dynamic programming can be used to find an optimal correspondence (eg., [48, 49, 56, 35]). Recently, more efficient shortest path algorithms have been used on related problems by Geiger et al.=-=[17]-=-, which we use in our implementation. Arkin et al.[2] use a more complex, geometrical algorithm specific to their formulation of the problem. Also popular have been various methods based on gradient d... |

155 |
The measurement of visual motion
- Hildreth
- 1983
(Show Context)
Citation Context ...arping or gradient descent for computing the amount of deformation required. There has been a great deal of work on this approach. For example, it has been applied to tracking non-rigid objects (eg., =-=[14, 19, 8, 24]-=-) where the amount of deformation between frames may be assumed to be relatively small. This approach has also been widely applied to handwriting recognition (eg., [30, 38, 33]) using measures of simi... |

154 | Learning flexible models from image sequences - Baumberg, Hogg - 1994 |

148 | Forms: a flexible object recognition and modeling system
- Zhu, Yuille
- 1996
(Show Context)
Citation Context ...function, and make similarity judgements based on semantics (eg. [37, 16, 29, 28]). Also relevant to our work, [26] deform shapes by aligning the principal modes of their mass and stiffness matrices, =-=[40]-=- combine part and deformable template methods in a novel way quite different from ours, [18] compare shapes by minimizing the Hausdorff measure between them, and [35] proposes aligning shapes by apply... |

128 |
Structural image restoration through deformable templates
- Amit, Grenander, et al.
- 1991
(Show Context)
Citation Context ...lting. Related work compares shapes by measuring the deformations needed to relate the 4 regions bounded by the shapes' contours, rather than the contours themselves (eg., Amit, Grenander and Piccioni=-=[1]-=- and Jain, Zhong, and Lakshmanan[28]). In Section 7 we briefly discuss some of the advantages and disadvantages of comparisons based on regions rather than the contours that bound them. Another approa... |

116 |
Recognizing and locating partially visible objects: The local feature-focus method
- Bolles, Cain
- 1982
(Show Context)
Citation Context ...s with a list of properties (features). These properties may be global (e.g., "the object is polygonal") or local. When global properties are used (see reviews in Duda and Hart[16] and Bolle=-=s and Cain[9]-=-) objects are represented as points in feature space. This representation is obtained by arranging the list of properties of an object as the components of a vector associated with the object. Similar... |

95 |
Tracking nonrigid objects in complex scenes
- Huttenlocher, Noh, et al.
- 1993
(Show Context)
Citation Context ...evant to our work, [26] deform shapes by aligning the principal modes of their mass and stiffness matrices, [40] combine part and deformable template methods in a novel way quite different from ours, =-=[18]-=- compare shapes by minimizing the Hausdorff measure between them, and [35] proposes aligning shapes by applying piecewise affine transformations. See [3] for further discussion of this work. Our appro... |

95 |
Recognition by Parts. The
- Pentland
- 1987
(Show Context)
Citation Context |

79 | Tracking points on deformable objects using curvature information,” ECCV
- Cohen, Ayache, et al.
- 1992
(Show Context)
Citation Context ... "stretching" of a contour. Hildreth[22] proposed computing motion flow between two contours with a cost function that penalizes for the variation in the magnitude of motion. Cohen, Ayache, =-=and Sulger[12]-=- propose matching contours extracted from moving images of biological structures (such as beating hearts) with a cost function based on the difference in curvature between the two curves. Tagare, O'Sh... |

68 |
Three dimensional object representation revisited
- Bajcsy, Solina
- 1987
(Show Context)
Citation Context |

67 | Learning physical description from functional descriptions, examples, and precedents
- Winston, Binford, et al.
- 1983
(Show Context)
Citation Context ...ilarity (eg. see [27], and a review in [10]). A third class of methods attempts to assign "semantic interpretation" to shapes, such as function, and make similarity judgements based on seman=-=tics (eg. [37, 16, 29, 28]-=-). Also relevant to our work, [26] deform shapes by aligning the principal modes of their mass and stiffness matrices, [40] combine part and deformable template methods in a novel way quite different ... |

64 |
Mathematical theories of shape: Do they model perception
- Mumford
- 1991
(Show Context)
Citation Context ...ognition; our work will attempt to combine these two approaches. We will more briefly describe other approaches to shape similarity, including feature based methods and functional methods. See Mumford=-=[37]-=- for another overview of methods of judging similarity. 2.1 Methods that Measure Deformations Elastic matching methods measure the difference between two shapes based on a specific correspondence betw... |

60 |
Generating and generalizing models of visual objects
- Connell, Brady
- 1987
(Show Context)
Citation Context |

60 |
Acheiving generalized object recognition through reasoning about association of function to structure
- Stark, Bowyer
- 1991
(Show Context)
Citation Context ...ilarity (eg. see [27], and a review in [10]). A third class of methods attempts to assign "semantic interpretation" to shapes, such as function, and make similarity judgements based on seman=-=tics (eg. [37, 16, 29, 28]-=-). Also relevant to our work, [26] deform shapes by aligning the principal modes of their mass and stiffness matrices, [40] combine part and deformable template methods in a novel way quite different ... |

60 |
A shock grammar for recognition
- Siddiqi, Kimia
- 1996
(Show Context)
Citation Context ...f shape. A number of other proposals have also been made for representing objects in terms of their possible decompositions into parts and the relationships between these parts (eg., Siddiqi and Kimia=-=[44]-=-, Kupeev and Wolfson[32], Zhu and Yuille[58]). Biederman[6] has produced a good deal of evidence for the contention that human object recognition is sensitive to the part structure of objects (see als... |

59 | Recognition by functional parts
- Rivlin, Dickinson, et al.
- 1995
(Show Context)
Citation Context ...thods that use function have attracted researchers from both fields of computer vision and artificial intelligence (Winston et al.[55], Ho[24], Stark and Bowyer[46] and Rivlin, Dickenson and Rosenfeld=-=[42]-=-). The replacement of geometric structure with semantics is potentially powerful. In particular, eccentrically designed objects may be recognized using this method even when they differ visually from ... |

58 | Adaptive elastic models for handprinted character recognition
- Hinton, Williams, et al.
- 1992
(Show Context)
Citation Context ... [32] suggest such an approach to judging the similarity of objects; [1] also propose such a method of comparing shape, with particular attention to producing fast algorithms that yield a metric; and =-=[22, 4, 15]-=- suggest learning a deformable model from examples. Work based on measuring deformations has not considered the role of part structure and articulation in shape similarity. Most of the above systems, ... |

55 |
The shape of smooth objects and the way contours end
- Koenderink, Doorn
- 1982
(Show Context)
Citation Context ...generally are defined to be convex or nearly convex shapes separated from the rest of the object at concavity extrema, as in Hoffman and Richards[25], or at inflections, as in Koenderink and van Doorn=-=[30]-=-. Many methods have been suggested for providing geometric descriptions of these parts, such as generalized cylinders (Binford[7]) and superquadrics (Bajcsy and Solina[3], Pentland[39]). Some methods ... |

50 |
Dynamic Planar Warping for Optical Character Recognition
- Levin, Pieraccini
- 1992
(Show Context)
Citation Context ...s of registering images of Kanji by using dynamic programming to register a 1-D projection of the 2-D image. The cost function is based on features of 1-D slices of the 2-D image. Levin and Pieraccini=-=[33]-=- also describe a method of warping 2-D images by registering 1-D slices of them, using dynamic programming. Our focus in this paper, however, is on how similarity is encoded in the cost function rathe... |

45 |
Cursive script recognition by elastic matching
- Tappert
- 1982
(Show Context)
Citation Context ...-rigid objects (eg., [14, 19, 8, 24]) where the amount of deformation between frames may be assumed to be relatively small. This approach has also been widely applied to handwriting recognition (eg., =-=[30, 38, 33]-=-) using measures of similarity that tend to be particularly crafted to that domain. Deformable templates that reflect the specific shape of known objects have also been applied to contour extraction a... |

44 |
Elastic Matching of Line Drawings
- Burr
- 1981
(Show Context)
Citation Context ...nce between corresponding points along the two contours. Tappert[48] compares the difference between the angle and height of corresponding points, also for character recognition. In related work, Burr=-=[11]-=- compares contours based on the distance and angle between the contours. Related approaches are also taken by Tsai and Yu[49], and Mehrota and Grosky[36], the latter focusing on the problem of indexin... |

35 | Learning object recognition models from images
- Pope, Lowe
- 1993
(Show Context)
Citation Context ...s are discussed in [7, 12, 13]. Another approach to shape comparison describes shapes with a list of properties and their relations, using pattern matching techniques to determine similarity (eg. see =-=[27], and a re-=-view in [10]). A third class of methods attempts to assign "semantic interpretation" to shapes, such as function, and make similarity judgements based on semantics (eg. [37, 16, 29, 28]). Al... |

29 |
Partial shape recognition using dynamic programming
- Gorman, Mitchell, et al.
- 1988
(Show Context)
Citation Context ...main unmatched, under the assumption that the true corresponding portions of the contour are not visible. We now consider how a proposed cost function should deal with an hypothesized occlusion. (See =-=[11]-=- for another approach to this problem). The cost function for occlusions should reflect three factors. First, it is important to take account of the similarity between the relative positions of the be... |

28 |
Attributed string matching with merging for shape recognition
- Tsai, Yu
- 1985
(Show Context)
Citation Context ...d to that domain. Deformable templates that reflect the specific shape of known objects have also been applied to contour extraction and modelling ([39, 31, 21]). Of particular relevance to our work: =-=[32]-=- suggest such an approach to judging the similarity of objects; [1] also propose such a method of comparing shape, with particular attention to producing fast algorithms that yield a metric; and [22, ... |

17 | A generic system for classifying variable objects using flexible template matching
- Lanitis, Taylor, et al.
- 1993
(Show Context)
Citation Context ... [32] suggest such an approach to judging the similarity of objects; [1] also propose such a method of comparing shape, with particular attention to producing fast algorithms that yield a metric; and =-=[22, 4, 15]-=- suggest learning a deformable model from examples. Work based on measuring deformations has not considered the role of part structure and articulation in shape similarity. Most of the above systems, ... |

17 |
Shapes, shocks, and deformations
- Kimia, Tannenbaum, et al.
- 1995
(Show Context)
Citation Context ...is is clearly not desirable and our approach overcomes this problem. Finally, many methods of curve description have been proposed that are based on curve evolution. For example, recently Kimia et al.=-=[29, 44]-=- have proposed curve evolution methods based on reaction-diffusion equations, using the results of this evolution to describe the shape of an object. It is natural to consider a diffusion process to m... |

17 | On Shape Similarity
- Kupeev, Wolfson
- 1999
(Show Context)
Citation Context ...er proposals have also been made for representing objects in terms of their possible decompositions into parts and the relationships between these parts (eg., Siddiqi and Kimia[44], Kupeev and Wolfson=-=[32]-=-, Zhu and Yuille[58]). Biederman[6] has produced a good deal of evidence for the contention that human object recognition is sensitive to the part structure of objects (see also Siddiqi, Tresness, 6 F... |

17 |
Shape matching utilizing indexed hypotheses generation and testing
- Mehrotra, Grosky
- 1989
(Show Context)
Citation Context ...or character recognition. In related work, Burr[11] compares contours based on the distance and angle between the contours. Related approaches are also taken by Tsai and Yu[49], and Mehrota and Grosky=-=[36]-=-, the latter focusing on the problem of indexing into a data base of shapes. Similar approaches have been used in work on motion tracking. In some cases this work also incorporates a cost that measure... |

16 |
Tracking Non-Rigid Objects
- Huttenlocher, Noh, et al.
- 1993
(Show Context)
Citation Context .... Pentland and Sclaroff[40] construct mass and stiffness matrices for the given shapes, and deform the shapes by aligning the principal modes of their mass and stiffness matrices. Huttenlocher et al.,=-=[26, 27]-=- compare two shapes by applying the rigid (or affine) transformation that minimizes the Hausdorf measure between them. To allow for occlusion only a certain fixed fraction of the contour points are us... |