## Shock Graphs and Shape Matching (1998)

### Cached

### Download Links

- [www.cs.yale.edu]
- [www.cim.mcgill.ca]
- [www.cs.rutgers.edu]
- [www.cim.mcgill.ca]
- [www.cs.toronto.edu]
- [www.cim.mcgill.ca]
- [www.cim.mcgill.edu]
- DBLP

### Other Repositories/Bibliography

Citations: | 223 - 35 self |

### BibTeX

@MISC{Siddiqi98shockgraphs,

author = {Kaleem Siddiqi and Ali Shokoufandeh and Sven J. Dickinson and Steven W. Zucker},

title = {Shock Graphs and Shape Matching},

year = {1998}

}

### Years of Citing Articles

### OpenURL

### Abstract

We have been developing a theory for the generic representation of 2-D shape, where structural descriptions are derived from the shocks (singularities) of a curve evolution process, acting on bounding contours. We now apply the theory to the problem of shape matching. The shocks are organized into a directed, acyclic shock graph, and complexity is managed by attending to the most significant (central) shape components first. The space of all such graphs is highly structured and can be characterized by the rules of a shock graph grammar. The grammar permits a reduction of a shock graph to a unique rooted shock tree. We introduce a novel tree matching algorithm which finds the best set of corresponding nodes between two shock trees in polynomial time. Using a diverse database of shapes, we demonstrate our system's performance under articulation, occlusion, and changes in viewpoint. Keywords: shape representation; shape matching; shock graph; shock graph grammar; subgraph isomorphism. 1 I...

### Citations

11365 |
Computers and Intractability - A Guide to the Theory of NPCompletness. Freeman and Co
- Garey, Johnson
- 1979
(Show Context)
Citation Context ...d H = (V 2;E 2), nd the maximum integer k, such that there exists two subsets of cardinality k, E 0 1 E 1 and E 0 2 E 2, and the induced subgraphs G 0 =(V 1;E 0 1) and H 0 =(V 2;E 0 2) are isomorphic =-=[17]-=-. Further, since our shock graphs are labeled graphs, consistency between node labels must be enforced in the isomorphism. The largest subgraph isomorphism problem, can be formulated as af0; 1g intege... |

2982 | Eigenfaces for Recognition
- Turk, Pentland
- 1991
(Show Context)
Citation Context ...e again they are not easily applied to natural objects. Second, a number of appearance-based methods have emerged which essentially treat the raw image as a single feature in a high-dimensional space =-=[54, 34]-=-. Whereas such techniques might succeed at recognizing particulars of a speci c class, e.g., faces, they cannot predict entry-level categories because there is no abstraction from image data to a mode... |

1204 |
Image Analysis and Mathematical Morphology
- Serra, Ed
- 1982
(Show Context)
Citation Context ...(x; ) and N(x; )6=;and N(x; )isnot connected; and 4. type 1 otherwise. It should be clear that there is a relationship between the above coloring and the velocity function dR dx along the medial axis =-=[48]-=-. In separate work, we are investigating the Morse function properties of the signed distance function, and the Arnold classi cation of singularities [3]. In Figure 7 we provide numerical examples of ... |

1000 |
Visual learning and recognition of 3d objects from appearance
- Murase, Nayar
- 1995
(Show Context)
Citation Context ...e again they are not easily applied to natural objects. Second, a number of appearance-based methods have emerged which essentially treat the raw image as a single feature in a high-dimensional space =-=[54, 34]-=-. Whereas such techniques might succeed at recognizing particulars of a speci c class, e.g., faces, they cannot predict entry-level categories because there is no abstraction from image data to a mode... |

953 | Sethian Fronts Propagating with CurvatureDependent Speed
- Osher, A
(Show Context)
Citation Context ...sed a linear di usion equation which can be used to build skeletal descriptions directly from greyscale images [52], and is computationally more e cient than those based on standard level set methods =-=[36]-=-. It leads to an approximation of the reaction di usion space introduced in [23]. Tek et al. have used an orientation propagating distance function to extract symmetries from fragmented contours, labe... |

660 | Shape modeling with front propagation: a level set approach
- Malladi, Sethian, et al.
- 1995
(Show Context)
Citation Context ...r planar shape recognition [14], obtained from an axial decomposition introduced in [42]. Among the many applications of curve evolution to problems in computer vision and image processing, e.g., see =-=[2, 45, 30]-=-, only a handful have addressed the problem of shape representation. Tari et. al have proposed a linear di usion equation which can be used to build skeletal descriptions directly from greyscale image... |

590 | Basic objects in natural categories
- Rosch, Mervis, et al.
- 1976
(Show Context)
Citation Context ...s the presence of a particular object, such as a dog, before realizing it is either a Siberian Husky or that it is \Loki", a particular Siberian. This example, modi ed from important studies by Rosch =-=[43]-=-, suggests that there is an organization to our object memory, and that this organization facilitates recognition. Initially, particular instances are not recognized; rather, objects are rst categoriz... |

563 |
A n5=2 algorithm for maximum matchings in bipartite graphs
- Hopcroft, Karp
- 1973
(Show Context)
Citation Context ...umber of vertices in the two graphs. The complexitywas further reduced by Reyner [40] to O(n 1:5 1 n 2) (assuming n 1 n2), through a reduction to the bipartite matching algorithm of Hopcraft and Karp =-=[21]-=-. If we could transform our directed acyclic shock graphs to nite rooted trees, we could pursue a polynomial time solution to our problem. In the following subsections we show that for any shock graph... |

499 | Interior point methods in semidefinite programming with applications to combinatorial optimization - Alizadeh - 1995 |

428 |
Recognition by linear combinations of models
- Ullman, Basri
- 1991
(Show Context)
Citation Context ...an several objects. In alignment schemes, the emphasis is not on boundary encoding, but on accounting for the di erences between an observed and a stored shape. A clever algorithm by Ullman and Basri =-=[56]-=- interpolates from a linear combination of 2-D views, and impressive results on a Volkswagen image were reported. However, to achieve these results, the edge maps were manually edited so that only tho... |

340 |
Image Selective Smoothing and Edge Detection by Nonlinear Diffusion
- Alvarez, Lions, et al.
- 1992
(Show Context)
Citation Context ...r planar shape recognition [14], obtained from an axial decomposition introduced in [42]. Among the many applications of curve evolution to problems in computer vision and image processing, e.g., see =-=[2, 45, 30]-=-, only a handful have addressed the problem of shape representation. Tari et. al have proposed a linear di usion equation which can be used to build skeletal descriptions directly from greyscale image... |

305 | Use of active shape models for locating structures in medical images - Cootes, Hill, et al. - 1994 |

299 | A graduated assignment algorithm for graph matching
- Gold, Rangarajan
- 1996
(Show Context)
Citation Context ...4], little progress was made. Research continues on graph isomorphism algorithms for vision applications, but examples are still typically based on graphs derived from feature points and image curves =-=[19]-=-. Probability measures have also been placed on images and image curves in an attempt to provide a priori information suitable to guide matching [20]. Image curves are also at the heart of boundary-ba... |

287 |
Elements of the Theory of Computation
- Lewis, Papadimitriou
- 1981
(Show Context)
Citation Context ...ng unions, and so on. It operates by beginning at the start symbol and repeatedly replacing the left-hand side of a rule by the corresponding right-hand side until no further replacements can be made =-=[28]-=-. It is the SGG that captures the beauty of shock graphs, because the rules embody constraints from the domain of curve evolution. In particular, Proposition 2 The rewrite rules of the SGG are su cien... |

227 |
vanDoorn, Internal representation of solid shape with respect to vision
- Koenderink, J
- 1979
(Show Context)
Citation Context ... the boundaries of objects implies a viewpoint dependency to shape recognition [8], but does not specify which features to use for each view. Aspect graphs were introduced by Koenderink and van Doorn =-=[25]-=- to enumerate topologically-distinct views [26] 2svia singular or catastrophic events. For example, when a cup is rotated, there is a particular viewpoint from which the handle just becomes visible; t... |

225 |
General Pattern Theory
- Grenander
- 1994
(Show Context)
Citation Context ...raphs derived from feature points and image curves [19]. Probability measures have also been placed on images and image curves in an attempt to provide a priori information suitable to guide matching =-=[20]-=-. Image curves are also at the heart of boundary-based descriptions, such as those of Ho - man and Richards [41], and alignment techniques, such as the one proposed by Ullman [5]. However, the Ho man ... |

218 |
Syntactic Pattern Recognition and Applications
- Fu
- 1982
(Show Context)
Citation Context ...entry-level, generic descriptions. 1s1.1 Classical Aspects of Shape Recognition Rosch's experimental observation that basic-level descriptions precede particulars was made about the same time that Fu =-=[15]-=- and others were introducing syntactic pattern recognition. Fu's goal was to de ne a grammar for patterns, and then to specify automata that could recognize this grammar. However, this program lost fa... |

217 | A theory of multi-scale, curvature-based shape representation for planar curves - Mokhtarian, Mackworth - 1992 |

187 | Modal Matching for Correspondence and Recognition - Sclaroff, Pentland - 1995 |

179 | Matching hierarchical structures using association graphs - Pelillo, Siddiqi, et al. - 1407 |

154 |
Biological shape and visual science
- Blum
- 1973
(Show Context)
Citation Context ...photographs of animals, but would fail at separating, for example, photographs of horses from photographs of hands. In important contrast to the boundary based techniques discussed earlier was Blum's =-=[6]-=- medial axis transform|or skeleton|which preceded Rosch by about a decade. Blum's skeleton is area-based, and provides a description of shapes via the loci of centers of covering balls. The skeleton h... |

152 | FORMS: a flexible object recognition and modeling system - Zhu, Yuille - 1996 |

148 | Psychophysical support for a two-dimensional view interpolation theory of object recognition - Bülthoff, Edelman - 1992 |

145 |
3-d shape recovery using distributed aspect matching
- Dickinson, Pentland, et al.
- 1992
(Show Context)
Citation Context ...sion approaches to view-based modeling fall broadly into two classes. First, there are feature-based methods which represent each view as a collection of line segments, curves, corners, regions, etc. =-=[22, 10, 11, 39]-=-. The success of such methods depends largely on the extent to which the features are present and can be reliably extracted; once again they are not easily applied to natural objects. Second, a number... |

139 | Segmentation, registration, and measurement of shape variation via image object shape - Pizer, Fritsch, et al. - 1996 |

137 |
Invariant descriptors for 3d object recognition and pose
- Forsyth, Mundy, et al.
- 1991
(Show Context)
Citation Context ...e; thereafter, only geometric variations take place until, at another singular viewpoint, the handle disappears. However, aspect graphs and other methods based on algebraic and di erential invariants =-=[13]-=- were successfully de ned only for speci c classes of algebraic surfaces that t only few (man-made) objects. The techniques are typically di cult to extend to natural objects. Computer vision approach... |

131 | Discrete Voronoi Skeletons
- Ogniewicz
- 1993
(Show Context)
Citation Context ...ot automatically, and as with the Fu and the Ullman approaches, the features have to be edited to provide a basis for matching. One option that is worth stressing is the use of hierarchical skeletons =-=[35]-=-, because it attempts to capture a notion of \scale" for objects. This is important because, should such scales be available, coarse-to- ne matching strategies could be employed; see also Burbeck and ... |

119 | Smoothed local symmetries and their implementation - Brady, Asada - 1984 |

108 | Determining the similarity of deformable shapes
- Basri, Costa, et al.
- 1995
(Show Context)
Citation Context ...tions, with some robustness to occlusion [33, 38]. Basri et al. have proposed various models for measuring the cost of deforming one contour into another, while taking into account its part structure =-=[4]-=-. Gdalyahu and Weinshall have also proposed metric functions for measuring the similarity between two closed planar curves [18]. However, the previous three methods do not explicitly account for a sha... |

93 |
On recognizing and positioning curved 3D objects from image contours
- KRIEGMAN
- 1990
(Show Context)
Citation Context ...dependency to shape recognition [8], but does not specify which features to use for each view. Aspect graphs were introduced by Koenderink and van Doorn [25] to enumerate topologically-distinct views =-=[26]-=- 2svia singular or catastrophic events. For example, when a cup is rotated, there is a particular viewpoint from which the handle just becomes visible; thereafter, only geometric variations take place... |

83 | Affine invariant scale-space - Sapiro, Tannenbaum - 1993 |

82 | Symmetry-based indexing of image databases - Sharvit, Chan, et al. - 1998 |

74 |
The Graph Isomorphism Problem: Its Structural Complexity
- Köbler, Schöning, et al.
- 1993
(Show Context)
Citation Context ...ted acyclic graph. This has important consequences for object matching because the problem of searching directed acyclic graphs is computationally much simpler than that of searching arbitrary graphs =-=[24]-=-. Second, since there exist rules in the SGG whose left-hand sides do not consist of single nonterminals, the SGG is not context-free. Third, the rewrite rules indicate that a 2-shock and a 4-shock ca... |

74 | Area and length minimizing flows for shape segmentation - Siddiqi, Lauziere, et al. - 1998 |

73 |
A process grammar for shape
- Leyton
- 1988
(Show Context)
Citation Context ...ined because the events that govern the birth, combination, and death of shock groups can be abstracted into a small number of rewrite rules, shown in Figure 2. In analogy to Leyton's Process Grammar =-=[29]-=-, the rules have been grouped according to the semantic processes that they characterize, although the alphabet of shock types that they operate on is quite di erent from boundary-based codons. De nit... |

72 | Object representation by cores: Identifying and representing primitive spatial regions
- Burbeck, Pizer
- 1995
(Show Context)
Citation Context ...use it attempts to capture a notion of \scale" for objects. This is important because, should such scales be available, coarse-to- ne matching strategies could be employed; see also Burbeck and Pizer =-=[9]-=-. In recentwork, Sclaro and Pentland have addressed the problem of 2-D shape matching 3susing a modal representation corresponding to a shape's generalized axes of symmetry [47]. This compact represen... |

66 | shocks, and deformations I: The components of two-dimensional shape and the reaction-diffusion space,” Intern - Kimia, Tannenbaum, et al. - 1995 |

63 | Optimality conditions and duality theory for minimizing sums of the largest eigenvalues of symmetric matrices
- OVERTON, WOMERSLEY
- 1993
(Show Context)
Citation Context ...ct, i.e., U V = X X Ui;jVi;j. For any square matrix U, wede ne trace(U) = P i Ui;i. i j Let I denote the identity matrix having suitable dimensions. The following result, due to Overton and Womersley =-=[37]-=-, characterizes the sum of the rst k largest eigenvalues of a symmetric matrix in the form of a semide nite convex programming problem: Theorem 1 For the sum of the rst k eigenvalues of a symmetric ma... |

62 |
Evolutions equations for continuous-scale morphology
- Brockett, Maragos
- 1992
(Show Context)
Citation Context ...n. The constant 0 controls the regularizing e ects of curvature . When is large, the equation becomes a geometric heat equation; when = 0, the equation is equivalent to Blum's grass re transformation =-=[7, 23]-=-. In this paper, we shall only be interested in the latter case, under which the evolution equation is hyperbolic and shocks [27], or entropy-satisfying singularities can form. Here we shall ignore th... |

61 |
A shock grammar for recognition
- Siddiqi, Kimia
- 1996
(Show Context)
Citation Context ...emantically equivalent to a ~3 in a speci c context. The SG's for a variety of shapes are shown in Figure 8. All the graphs were generated automatically from the output of the shock detection process =-=[49]-=- displayed in Figure 7. Following the third observation, only label types ~1 and ~3 have been explicitly assigned. A ~3 with a parent ~1 ateach end acts as a 2 (a neck), and a ~3 with a # as a parent ... |

60 |
Hierarchical decomposition and axial shape description
- ROM, MEDIONI
- 1993
(Show Context)
Citation Context ...re global properties such as symmetry. Finally,Francois and Medioni have proposed a connection hierarchy of parts for planar shape recognition [14], obtained from an axial decomposition introduced in =-=[42]-=-. Among the many applications of curve evolution to problems in computer vision and image processing, e.g., see [2, 45, 30], only a handful have addressed the problem of shape representation. Tari et.... |

51 | An efficient approximation algorithm for the survivable network design problem
- Gabow, Goemans, et al.
(Show Context)
Citation Context ...er, and has the value1otherwise. Next, we form a bipartite edge weighted graph G(V 1;V 2;E G) with edge weights from the matrix (G; H). 6 Using the scaling algorithm of Goemans, Gabow, and Williamson =-=[16]-=-, we then nd the maximum cardinality, minimum weight matching inG. This results in a list of node correspondences between G and H, calledM 1, that can be ranked in decreasing order of similarity. From... |

51 |
Automatic generation of object recognition programs
- Ikeuchi, Kanade
- 1988
(Show Context)
Citation Context ...sion approaches to view-based modeling fall broadly into two classes. First, there are feature-based methods which represent each view as a collection of line segments, curves, corners, regions, etc. =-=[22, 10, 11, 39]-=-. The success of such methods depends largely on the extent to which the features are present and can be reliably extracted; once again they are not easily applied to natural objects. Second, a number... |

45 | The alignment of objects with smooth surfaces
- Basri, Ullman
- 1993
(Show Context)
Citation Context ...le to guide matching [20]. Image curves are also at the heart of boundary-based descriptions, such as those of Ho - man and Richards [41], and alignment techniques, such as the one proposed by Ullman =-=[5]-=-. However, the Ho man and Richards \codon" vocabulary is only an intermediate step toward more abstract part descriptions, and remains to be completed. It is an attempt to restrict the graph represent... |

44 |
An analysis of a good algorithm for the subtree problem, corrected
- Verma, Reyner
- 1989
(Show Context)
Citation Context ... describe once such technique, involving the solution of 2n 1n 2 network ow problems, where n 1 and n 2 represent the number of vertices in the two graphs. The complexitywas further reduced by Reyner =-=[40]-=- to O(n 1:5 1 n 2) (assuming n 1 n2), through a reduction to the bipartite matching algorithm of Hopcraft and Karp [21]. If we could transform our directed acyclic shock graphs to nite rooted trees, w... |

44 | Deformable Prototypes for Encoding Shape Categories in Image Databases - Sclaroff - 1997 |

42 | Codon constraints on closed 2D shapes
- Richards, Hoffman
- 1985
(Show Context)
Citation Context ...d image curves in an attempt to provide a priori information suitable to guide matching [20]. Image curves are also at the heart of boundary-based descriptions, such as those of Ho - man and Richards =-=[41]-=-, and alignment techniques, such as the one proposed by Ullman [5]. However, the Ho man and Richards \codon" vocabulary is only an intermediate step toward more abstract part descriptions, and remains... |

37 |
Model matching for correspondence and recognition
- Sclaroff, Pentland
- 1995
(Show Context)
Citation Context ... also Burbeck and Pizer [9]. In recentwork, Sclaro and Pentland have addressed the problem of 2-D shape matching 3susing a modal representation corresponding to a shape's generalized axes of symmetry =-=[47]-=-. This compact representation has been used for indexing [46], and o ers a frequency-like (coarse to ne) decomposition of a shape. However, its global nature makes it sensitive to large occlusion. Zhu... |

37 | Representation of self-similarity of shapes - Geiger, Liu, et al. - 2003 |

36 | Learning Object Recognition Models from Images
- Pope, Lowe
- 1993
(Show Context)
Citation Context ...sion approaches to view-based modeling fall broadly into two classes. First, there are feature-based methods which represent each view as a collection of line segments, curves, corners, regions, etc. =-=[22, 10, 11, 39]-=-. The success of such methods depends largely on the extent to which the features are present and can be reliably extracted; once again they are not easily applied to natural objects. Second, a number... |