## A Framework for Constructing Probability Distributions on the Space of Image Segmentations (1995)

Venue: | Computer Vision and Image Understanding |

Citations: | 6 - 3 self |

### BibTeX

@ARTICLE{Lavalle95aframework,

author = {Steven M. Lavalle and Seth A. Hutchinson},

title = {A Framework for Constructing Probability Distributions on the Space of Image Segmentations},

journal = {Computer Vision and Image Understanding},

year = {1995},

volume = {61},

pages = {203--230}

}

### OpenURL

### Abstract

The goal of traditional probabilistic approaches to image segmentation has been to derive a single, optimal segmentation, given statistical models for the image formation process. In this paper, we describe a new probabilistic approach to segmentation, in which the goal is to derive a set of plausible segmentation hypotheses and their corresponding probabilities. Because the space of possible image segmentations is too large to represent explicitly, we present a representation scheme that allows the implicit representation of large sets of segmentation hypotheses that have low probability. We then derive a probabilistic mechanism for applying Bayesian, model-based evidence to guide the construction of this representation. One key to our approach is a general Bayesian method for determining the posterior probability that the union of regions is homogeneous, given that the individual regions are homogeneous. This method does not rely on estimation, and properly treats the issu...

### Citations

7441 |
Probabilistic Reasoning in Intelligent Systems
- Pearl
- 1988
(Show Context)
Citation Context ...he case, it is possible to develop more general models for which P ((I ae ; E ae )) is affected 22 by considering evidence associated with R ae . For example, taxonomic hierarchies, analyzed by Pearl =-=[51]-=-, allow this to occur. Pearl gives an efficient method for propagating evidencebased, posterior probabilities throughout a hierarchy of events, but the construction of the hierarchy is not considered.... |

4363 |
Classification and Regression Trees
- Breiman, Friedman, et al.
- 1984
(Show Context)
Citation Context ...o these events as ground segment events, since such events can not be refined. There is a correspondence between generating a sequence of T -covers, and generating classification and regression trees =-=[10]-=-,[12],[26]. Classification and regression trees are used to represent a sample space efficiently, often for the purpose of pattern recognition. In terms of classification and regression trees, T -refi... |

4142 |
Pattern Classification and Scene Analysis
- Duda, Hart
(Show Context)
Citation Context ...ces. Bell combines clustering with a Monte-Carlo approach to segment radiograph images, to determine manufacturing defects [4]. Four basic components involved in most clustering algorithms are (e.g., =-=[20]-=-,[36]): (1) Define a feature metric space. (2) Determine feature values corresponding to pixels or regions. (3) Iteratively group pixels or regions with close features in the metric space (4) Terminat... |

4020 |
Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ... explicit dependency on some local neighborhood consisting of other random variables. The general model that is most often used in computer vision was introduced in a seminal paper by Geman and Geman =-=[27]-=-, in the context of image restoration. The appeal of the approach is the fact that any MRF formulation (which applies to a variety of image models) can be expressed as an energy minimization problem, ... |

2428 |
A mathematical theory of evidence
- Shafer
- 1976
(Show Context)
Citation Context ...that form a partition of \Pi, with each set in the partition corresponding to some segment from T i . The relationship between \Theta i and \Pi resembles the refinement relationship defined by Shafer =-=[57]-=-. In fact, the mapping f i is very similar to what Shafer terms a refining mapping; however, this should not be confused with our use of refinement in the context of creating new approximations for T ... |

2303 |
Algorithms for Clustering Data
- Jain, Dubes
- 1988
(Show Context)
Citation Context ...Bell combines clustering with a Monte-Carlo approach to segment radiograph images, to determine manufacturing defects [4]. Four basic components involved in most clustering algorithms are (e.g., [20],=-=[36]-=-): (1) Define a feature metric space. (2) Determine feature values corresponding to pixels or regions. (3) Iteratively group pixels or regions with close features in the metric space (4) Terminate bas... |

1483 |
Robot Vision
- Horn
- 1986
(Show Context)
Citation Context ...art, to its underconstrained nature. For example, Horn asserts that one of the primary difficulties in evaluating a segmentation method is the lack of a clear definition of the "correct" seg=-=mentation [32]-=-; and Szeliski argues that low-level image models often underconstrain the solution, and advocates the use of uncertainty estimation [63]. Jain and Binford assert that a key problem in vision research... |

988 | On the statistical analysis of dirty pictures
- Besag
- 1986
(Show Context)
Citation Context ...in a manner that guarantees convergence to the optimal energy state, but the rate of convergence can be slow in practice. Other techniques have been developed which yield performance tradeoffs. Besag =-=[7]-=- proposed the iterative conditional modes (ICM) method as a feasible alternative to stochastic relaxation. Marroquin et al. gave an approach, called maximizer of posterior marginals (MPM), which defin... |

694 |
Pattern Recognition: A Statistical Approach
- Devijver, Kittler
- 1982
(Show Context)
Citation Context ...py, which is a function of a probability distribution. A clear discussion of the characteristics of an entropy measure is provided in [3]. Several alternative, entropy-based functions can be found in =-=[17]-=-,[60]. An entropy measure can be used, for example, to select between different models, or to decide to combine several models synergistically. Szeliski argues that a measure of uncertainty can be use... |

308 |
Image segmentation techniques
- Haralick, Shapiro
- 1985
(Show Context)
Citation Context ... in the usual way. A region, R, is some connected subset of the image. In practice, most region-based segmentation algorithms begin by partitioning the image into an initial set of regions, R 8 (e.g.,=-=[31]-=-,[50],[55],[58]). This provides a computational advantage (since there are not as many potential groupings of data points to consider), and also allows statistical models to be effectively exploited [... |

249 |
Modeling and segmentation of noisy and textured images using Gibbs random fields
- Derin, Elliott
- 1987
(Show Context)
Citation Context ...es with the approach are the computational complexity of the optimization, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture [14],=-=[16]-=-,[48], color-constancy [15], blurring [27],[39], boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmenta... |

245 |
Information Theory
- Ash
- 1990
(Show Context)
Citation Context ... directly quantified. One natural measure is the information entropy, which is a function of a probability distribution. A clear discussion of the characteristics of an entropy measure is provided in =-=[3]-=-. Several alternative, entropy-based functions can be found in [17],[60]. An entropy measure can be used, for example, to select between different models, or to decide to combine several models synerg... |

241 |
The representation, recognition, and locating of 3-D objects
- Faugeras, Hebert
- 1986
(Show Context)
Citation Context ...sed recognition of objects composed of polynomial surface patches, three segments may be sufficient to determine the position of an object. Faugeras and Hebert provide conditions for which this holds =-=[23]-=-. Each three-segment group corresponds to a set of segmentations. One advantage of our work is that it allows the consideration of such partial segmentations, and their corresponding probabilities. We... |

239 | Approximate Calculation of Multiple Integrals - Stroud - 1971 |

236 |
Estimation of planar curves, surfaces and nonplanar space curves defined by implicit equations, with applications to edge and range image segmentation
- Taubin
- 1991
(Show Context)
Citation Context ... polynomial model, applying to 3D surfaces. A more general formulation of implicit polynomial models, pertaining to curves and surfaces of arbitrary degree and dimension, has been developed by Taubin =-=[64]-=-. A polynomial can be considered as a linear combination of monomial basis functions. An implicit polynomial equation is represented as OE(\Delta; u) = N X j=1 u j x a j 1 x b j 2 x c j 3 = 0 (67) wit... |

197 |
Bayesian Modeling of Uncertainty in Low-Level Vision
- Szeliski
- 1989
(Show Context)
Citation Context ...s the lack of a clear definition of the "correct" segmentation [32]; and Szeliski argues that low-level image models often underconstrain the solution, and advocates the use of uncertainty e=-=stimation [63]-=-. Jain and Binford assert that a key problem in vision research is that the segmentation problem is often ignored or is assumed to have been solved [38]. In spite of these criticisms, most, if not all... |

149 |
Parallel and deterministic algorithms from MRFs: Surface reconstruction and integration
- Geiger, Girosi
- 1990
(Show Context)
Citation Context ...nimization problem, in which parallelism can be exploited. The primary difficulties with the approach are the computational complexity of the optimization, and the problem of MRF parameter estimation =-=[25]-=-. The approach has been applied to modeling noise processes and texture [14],[16],[48], color-constancy [15], blurring [27],[39], boundary modeling [28],[39], and locally-dependent nonlinear image tra... |

146 |
Artificial intelligence
- Winston
- 1977
(Show Context)
Citation Context ...est probability in \Theta i . 5.2 Constructing Approximate Representations of S We have developed two algorithms for deriving approximate representations for S. The first uses a beam-search algorithm =-=[66]-=-, and the second generates the n segmentations with the highest posterior probability. For both S algorithms, it is necessary, at various stages, to choose a new initial region, R i , to begin the con... |

139 |
Structural Pattern Recognition
- Pavlidis
- 1977
(Show Context)
Citation Context ...he usual way. A region, R, is some connected subset of the image. In practice, most region-based segmentation algorithms begin by partitioning the image into an initial set of regions, R 8 (e.g.,[31],=-=[50]-=-,[55],[58]). This provides a computational advantage (since there are not as many potential groupings of data points to consider), and also allows statistical models to be effectively exploited [58]. ... |

133 |
Boundary detection by constrained optimization
- Geman, Geman, et al.
- 1990
(Show Context)
Citation Context ...on, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture [14],[16],[48], color-constancy [15], blurring [27],[39], boundary modeling =-=[28]-=-,[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with MRF line processes [18],[37],[49]. In recen... |

127 |
Simultaneous parameter estimation and segmentation of Gibbs random fields using simulated annealing
- Lakshmanan, Derin
- 1989
(Show Context)
Citation Context ...rgy function and parameter estimation for 3-D surfaces through a single performance functional 6 [62]. An alternative adaptive estimation/optimization scheme has been proposed by Lakshmanan and Derin =-=[44]-=-. Manjunath and Chellappa [48] argue that for texture models, estimation on small windows and simple, nearest-neighbor clustering can be used as a starting point for the energy optimization to yield r... |

126 |
Unsupervised texture segmentation using markov random field models
- Manjunath, Chellappa
- 1991
(Show Context)
Citation Context ...th the approach are the computational complexity of the optimization, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture [14],[16],=-=[48]-=-, color-constancy [15], blurring [27],[39], boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation ... |

105 |
The theory and practice of Bayesian image labeling
- Chou, Brown
- 1990
(Show Context)
Citation Context ...n the image. Cohen and Cooper give a parallel, hierarchical algorithm for optimizing the energy function in a Gaussian MRF [14]. An alternative approach to the minimization is taken by Chou and Brown =-=[13]-=-. The site values are arranged in a hierarchy, in which various nodes represent subsets of the set of labelings that can be assigned. Using this representation, a highest confidence first (HCF) method... |

100 |
Robust clustering with applications in computer vision
- Jolion, Meer, et al.
- 1991
(Show Context)
Citation Context ...mber of classes, and recent work has been done specifically addressing the problem of determining the number of classes, known as cluster validation, in the context of image segmentation applications =-=[40]-=-,[67]. The Markov random field (MRF) approach models the image as a lattice of random variables, with each variable having explicit dependency on some local neighborhood consisting of other random var... |

71 |
Optimal partitioning for classification and regression trees
- Chou
- 1991
(Show Context)
Citation Context ...se events as ground segment events, since such events can not be refined. There is a correspondence between generating a sequence of T -covers, and generating classification and regression trees [10],=-=[12]-=-,[26]. Classification and regression trees are used to represent a sample space efficiently, often for the purpose of pattern recognition. In terms of classification and regression trees, T -refinemen... |

66 |
An iterative growing and pruning algorithm for classification tree design
- Gelfand, Ravishankar, et al.
- 1991
(Show Context)
Citation Context ...ents as ground segment events, since such events can not be refined. There is a correspondence between generating a sequence of T -covers, and generating classification and regression trees [10],[12],=-=[26]-=-. Classification and regression trees are used to represent a sample space efficiently, often for the purpose of pattern recognition. In terms of classification and regression trees, T -refinement cor... |

63 |
Simple parallel hierarchical and relaxation algorithms for segment ing noncausal markovian random fields
- Cohen, Cooper
- 1987
(Show Context)
Citation Context ...culties with the approach are the computational complexity of the optimization, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture =-=[14]-=-,[16],[48], color-constancy [15], blurring [27],[39], boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 seg... |

62 |
Bayes factors and choice criteria for linear models
- Smith, Spiegelhalter
- 1980
(Show Context)
Citation Context ...9) has appeared recently in work from the statistics literature, and is termed a Bayes factor. Smith and Speigelhalter used a similar ratio for model selection between nested linear parametric models =-=[59]-=-. Aitken has developed a Bayes factor for model comparison that conditions the prior model on the data [2]. Kass and Vaidyanathan present and discuss some asymptotic approximations and sensitivity to ... |

59 |
Intergration, inference, and management of spatial information using bayesian network: Perceptual organization
- Sarkar, Boyer
- 1993
(Show Context)
Citation Context ...nd Binford et al. have considered them for model-based object recognition applications [1],[8]. Sarkar and Boyer have proposed Bayesian networks for a hierarchical organization of perceptual features =-=[56]-=-. Since P ((I ae ; E ae )) is known inductively, only one of the terms on the right side of (24) is required, say P ((I ae [ fR ae g; E ae )). To assist in the application of model-based evidence, we ... |

47 |
Posterior Bayes factors
- Aitkin
- 1991
(Show Context)
Citation Context ...eigelhalter used a similar ratio for model selection between nested linear parametric models [59]. Aitken has developed a Bayes factor for model comparison that conditions the prior model on the data =-=[2]-=-. Kass and Vaidyanathan present and discuss some asymptotic approximations and sensitivity to varying priors of the Bayes factor [42]. Petit also discusses priors, but with concern for robustness with... |

43 |
On Optimally Combining Pieces of Information, with Application to Estimating 3-D Complex-Object Position from Range Data
- Bolle
- 1986
(Show Context)
Citation Context ...st need to express the density corresponding to the displacement of an observed point from a given surface. We use a probability model for range-scanning error, used and justified by Bolle and Cooper =-=[9]-=-, and also used by Taubin [64]. The model asserts that density, p(ffiju), of the displacement of an observed point from the surface, OE(x; u), is a Gaussian random variable with zero mean and some kno... |

36 |
Supervised Texture Image Segmentation Using Feature Smoothing and Probabilistic Relaxation Techniques
- Hsiao, Sawchuck
- 1989
(Show Context)
Citation Context ...often required determine these probabilities, as in [54]. A more recent discussion of relaxation labeling, applied to the problem of supervised and unsupervised texture classification can be found in =-=[33]-=-,[34]. 3 Representing Segment and Segmentation Probability Spaces In this section we describe an approximate representation scheme that allows for explicit representation of segmentation hypotheses wi... |

35 |
Image restoration using Gibbs priors: Boundary modeling, treatment of blurring, and selection of hyperparameter
- Johnson, Wong, et al.
- 1991
(Show Context)
Citation Context ...lexity of the optimization, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture [14],[16],[48], color-constancy [15], blurring [27],=-=[39]-=-, boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with MRF line processes ... |

34 |
Bayesian inference in modelbased machine vision
- Binford, Levitt, et al.
- 1987
(Show Context)
Citation Context ...sidered. Pearl's more general Bayesian networks have also been applied to computer vision problems. Agosta, and Binford et al. have considered them for model-based object recognition applications [1],=-=[8]-=-. Sarkar and Boyer have proposed Bayesian networks for a hierarchical organization of perceptual features [56]. Since P ((I ae ; E ae )) is known inductively, only one of the terms on the right side o... |

32 |
Approximate Bayes factors and orthogonal parameters, with application to testing equality of two binomial proportions
- Kass, Vaidyanathan
- 1992
(Show Context)
Citation Context ...or for model comparison that conditions the prior model on the data [2]. Kass and Vaidyanathan present and discuss some asymptotic approximations and sensitivity to varying priors of the Bayes factor =-=[42]-=-. Petit also discusses priors, but with concern for robustness with respect to outliers [52]. The Bayes factor has also been carefully studied for evidence evaluation in a forensic science context [6]... |

31 |
Bayesian clustering for unsupervised estimation of surface and texture models
- Silverman, Cooper
- 1993
(Show Context)
Citation Context ...MRF energy optimization, and probabilistic relaxation. In this section, we briefly review work in these areas. Clustering has been applied to a variety of image types and models. Silverman and Cooper =-=[58]-=- segment intensity images into regions that can be approximated by planar or quadric surfaces. Bell combines clustering with a Monte-Carlo approach to segment radiograph images, to determine manufactu... |

31 |
A modelfitting approach to cluster validation with application to stockastic model-based image segmentation
- Zhang, Modestino
- 1990
(Show Context)
Citation Context ... polynomial surfaces, as in [58]. The decisions involved in the third step depend on the particular clustering algorithm chosen, such as agglomerative clustering [20],[58] and K-means clustering [24],=-=[67]-=-. Most clustering algorithms require specification of the number of classes, and recent work has been done specifically addressing the problem of determining the number of classes, known as cluster va... |

30 |
Spectral classification using fuzzy membership models
- Kent, Mardia
- 1986
(Show Context)
Citation Context ...recisely represented by the right hand side of (25). Our membership probability can alternatively be considered as a definition of fuzzy membership, with respect to the segment that contains R i [19],=-=[43]-=-. 4.1.1 IE-independent and IE-dependent models The probability P ((fR i ; R ae g; ;) j (I ae ; E ae )) in (25) depends on R ae , I ae and E ae . If a model uses information from all of these regions, ... |

29 |
MRF Model-Based Algorithm for Image Segmentation
- Dubes, Jain, et al.
- 1990
(Show Context)
Citation Context ..., boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with MRF line processes =-=[18]-=-,[37],[49]. In recent years there has been considerable interest in improving MRF energy optimization algorithms. Geman and Geman used a simulated annealing approach (also called stochastic relaxation... |

28 |
Weight of evidence: a brief survey
- Good
- 1985
(Show Context)
Citation Context ...with respect to outliers [52]. The Bayes factor has also been carefully studied for evidence evaluation in a forensic science context [6],[11],[21],[22]. Other references to Bayes factors include [5],=-=[30]-=-,[41]. 4.3 Other cases and extensions In this section we briefly present some extensions to the result presented in the previous section. The expressions in Section 4.3.1 represent the more general IE... |

26 |
myopia and naivete in computer vision systems, CVGIp:hnge Undentmding,53, No 1.~~112-117
- Jain, Binford
- 1991
(Show Context)
Citation Context ...and advocates the use of uncertainty estimation [63]. Jain and Binford assert that a key problem in vision research is that the segmentation problem is often ignored or is assumed to have been solved =-=[38]-=-. In spite of these criticisms, most, if not all, previous approaches to segmentation have been aimed at deriving a single, optimal segmentation result for a given scene. This segmentation is often us... |

24 | Segmentation of 3-D range images using pyramidal data structures
- Sabata, Arman, et al.
- 1990
(Show Context)
Citation Context ...ual way. A region, R, is some connected subset of the image. In practice, most region-based segmentation algorithms begin by partitioning the image into an initial set of regions, R 8 (e.g.,[31],[50],=-=[55]-=-,[58]). This provides a computational advantage (since there are not as many potential groupings of data points to consider), and also allows statistical models to be effectively exploited [58]. The o... |

23 | A bayesian segmentation methodology for parametric image models - LaValle, Hutchinson - 1995 |

20 |
Bayesian testing of precise hypotheses (with discussion
- Berger, Delampady
- 1987
(Show Context)
Citation Context ...ess with respect to outliers [52]. The Bayes factor has also been carefully studied for evidence evaluation in a forensic science context [6],[11],[21],[22]. Other references to Bayes factors include =-=[5]-=-,[30],[41]. 4.3 Other cases and extensions In this section we briefly present some extensions to the result presented in the previous section. The expressions in Section 4.3.1 represent the more gener... |

17 |
Color image segmentation using Markov random fields
- Daily
- 1989
(Show Context)
Citation Context ...e computational complexity of the optimization, and the problem of MRF parameter estimation [25]. The approach has been applied to modeling noise processes and texture [14],[16],[48], color-constancy =-=[15]-=-, blurring [27],[39], boundary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with... |

16 |
Modeling sensors: Toward automatic generation of object recognition program
- Ikeuchi, Kanade
- 1989
(Show Context)
Citation Context ...le models that can be used. For different imaging systems, other models may be more appropriate. Ikeuchi and Kanade provide a detailed discussion of the modeling of a variety of range-imaging sensors =-=[35]-=-. Since taking the sum of squares of Gaussian densities yields the chi-square density, the degradation density, for unit variance, using (69) is p(y k ju k ) =s2 m k (y k ) = 1 2 m k =2 \Gamma(m k =2)... |

13 |
MRF Model-Based Segmentation of Range Images
- Jain, Nadabar
- 1990
(Show Context)
Citation Context ...ndary modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with MRF line processes [18],=-=[37]-=-,[49]. In recent years there has been considerable interest in improving MRF energy optimization algorithms. Geman and Geman used a simulated annealing approach (also called stochastic relaxation) to ... |

10 |
A general approach to parameter evaluation in fuzzy digital pictures
- Dubois, Jaulent
- 1987
(Show Context)
Citation Context ...een precisely represented by the right hand side of (25). Our membership probability can alternatively be considered as a definition of fuzzy membership, with respect to the segment that contains R i =-=[19]-=-,[43]. 4.1.1 IE-independent and IE-dependent models The probability P ((fR i ; R ae g; ;) j (I ae ; E ae )) in (25) depends on R ae , I ae and E ae . If a model uses information from all of these regi... |

9 |
Pinchin R: Statistical inference in crime investigations using deoxyribonucleic acid profiling
- Berry, Evett
- 1992
(Show Context)
Citation Context ...42]. Petit also discusses priors, but with concern for robustness with respect to outliers [52]. The Bayes factor has also been carefully studied for evidence evaluation in a forensic science context =-=[6]-=-,[11],[21],[22]. Other references to Bayes factors include [5],[30],[41]. 4.3 Other cases and extensions In this section we briefly present some extensions to the result presented in the previous sect... |

9 |
Parameter estimation in MRF line process models, in
- Nadabar, Jain
- 1992
(Show Context)
Citation Context ... modeling [28],[39], and locally-dependent nonlinear image transformations [27]. This model has also been considered for range image 5 segmentation by modeling edges with MRF line processes [18],[37],=-=[49]-=-. In recent years there has been considerable interest in improving MRF energy optimization algorithms. Geman and Geman used a simulated annealing approach (also called stochastic relaxation) to deter... |