## Computational Models of Perceptual Organization (2003)

### Cached

### Download Links

Venue: | Robotics Institute, Carnegie Mellon University |

Citations: | 5 - 0 self |

### BibTeX

@TECHREPORT{Yu03computationalmodels,

author = {Stella Yu},

title = {Computational Models of Perceptual Organization},

institution = {Robotics Institute, Carnegie Mellon University},

year = {2003}

}

### OpenURL

### Abstract

Perceptual organization refers to the process of organizing sensory input into coherent and interpretable perceptual structures. This process is challenging due to the chicken-and-egg nature between the various sub-processes such as image segmentation, figure-ground segregation and object recognition. Low-level processing requires the guidance of high-level knowledge to overcome noise; while high-level processing relies on low-level processes to reduce the computational complexity. Neither process can be sufficient on its own. Consequently, any system that carries out these processes in a sequence is bound to be brittle. An alternative system is one in which all processes interact with each other simultaneously. In this thesis, we develop a set of simple yet realistic interactive processing models for perceptual organization. We model the processing in the framework of spectral graph theory, with a criterion encoding the overall goodness of perceptual organization. We derive fast solutions for near-global optima of the criterion, and demonstrate the efficacy of the models on segmenting a wide range of real images. Through these models, we are able to capture a variety of perceptual phenomena: a unified treatment of various grouping, figure-ground and depth cues to produce popout, region segmentation and depth segregation in one step; and a unified framework for integrating bottom-up and top-down information to produce an object segmentation from spatial and object attention. We achieve these goals by empowering current spectral graph methods with a principled solution for multiclass spectral graph partitioning; expanded repertoire of grouping cues to include similarity, dissimilarity and ordering relationships; a theory for integrating sparse grouping cues; and a model ...

### Citations

4055 |
Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...t al., 1993; Geiger and Kumaran, 1996; Williams and Jacobs, 1997), region growing (Hong and Rosenfeld, 1984; Adams and Bischof, 1994), region competition (Zhu and Yuille, 1996), Markov random fields (=-=Geman and Geman, 1984-=-; Blake and Zisserman, 1987), level-set and variational methods (Sethian, 1996), graph cuts (Shi and Malik, 2000), etc. Figure 1.2: Poor imaging conditions and the interactions between foreground and ... |

3436 | A Computational Approach To Edge Detection
- Canny
- 1986
(Show Context)
Citation Context ...arlier attempts at image segmentation focused on simple low-level feature analysis and enhancement, giving rise to approaches such as thresholding (Sahoo et al., 1988) and edge linking (Witkin, 1983; =-=Canny, 1986-=-; Deriche, 1990). In a basic thresholding approach, a histogram of intensities for all the pixels in the image is first collected. Then a threshold is found by locating the deepest valley of the histo... |

3277 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...ing laws such as boundary smoothness need to be employed. Knowledge of such generic priors on boundary and/or region properties is used in various approaches of image segmentation: 2sactive contours (=-=Kass et al., 1988-=-; Cohen, 1991; Ronfard, 1994), boundary detection (Mumford and Shah, 1985; Nitzberg et al., 1993; Geiger and Kumaran, 1996; Williams and Jacobs, 1997), region growing (Hong and Rosenfeld, 1984; Adams ... |

2928 |
Introduction to Statistical Pattern Recognition, Electrical Science Series
- Fukunaga
- 1972
(Show Context)
Citation Context ...ead) of a person walking on a treadmill in all 32 images of a complete gait cycle. Using the magnitude thresholded edge orientations in the hand-labeled boxes as features, a linear Fisher classifier (=-=Fukunaga, 1990-=-) is trained for each body part. In order to account for the appearance changes of the limbs through the gait cycle, two separate models are used for each arm and each leg, bringing the total number o... |

2775 | Normalized cuts and image segmentation
- SHI, MALIK
- 1997
(Show Context)
Citation Context ...s, 1999; Ng 18 £ £ £set al., 2002), others use more eigenvectors than partitions required (Alpert et al., 1995; Malik et al., 2001). Sometimes, the first trivial eigenvector is discarded (Hall, 197=-=0; Shi and Malik, 1997-=-). Some works use eigenvectors literally as point coordinates (Shi and Malik, 2000; Alpert and Kahng, 1995b), while many normalize these points to have unit lengths to construct interpretable affinity... |

1725 | Nonlinear dimensionality reduction by locally linear embedding
- ROWEIS, SAUL
- 2000
(Show Context)
Citation Context ...generative model to explain data clusters, spectral graph partitioning methods (Chung, 1997) are an alternative that is more closely related to multi-dimensional scaling and locally linear embedding (=-=Roweis and Saul, 2000-=-), the goal of which is to preserve certain relationships (e.g. distance) among data points in a lower-dimensional representation. K-class spectral clustering is regarded as one that embeds data point... |

1485 | Fast approximate energy minimization via graph cuts
- BOYKOV, VEKSLER, et al.
- 2001
(Show Context)
Citation Context ...e particular forms of ε, such as the above mentioned probability criteria using generative models and minimum cuts criteria in discriminative approaches (Ishikawa and Geiger, 1998; Roy and Cox, 1998;=-= Boykov et al., 1999-=-), the constraints in Eqn (5.3) can be trivially incorporated in an algorithm that optimizes the objective. For the former, Markov Chain Monte Carlo is a general solution technique and the constraints... |

1186 | On spectral clustering: Analysis and an algorithm
- Ng, Jordan, et al.
- 2002
(Show Context)
Citation Context ...nvector components and are thus mapped to nearby n-dimensional points, with each coordinate specified by one of the n eigenvectors. Various clustering heuristics such as K-means (Shi and Malik, 2000; =-=Ng et al., 2002-=-), transportation (Barnes, 82), dynamic programming (Alpert and Kahng, 1995b), greedy pruning or exhaustive search (Shi and Malik, 2000) are subsequently employed on the new point sets to retrieve par... |

859 | Text classification from labeled and unlabeled documents using EM
- Nigam, McCallum, et al.
- 2000
(Show Context)
Citation Context ...th little extra cost. Our work is also closely linked to the transduction problem, the goal of which is to complete the labeling of a partially labeled dataset (Joachims, 1999; Jaakkola et al., 1999; =-=Nigam et al., 1999-=-; Szummer and Jaakkola, 2001). If the labeled data set is rich enough to characterize both the structures of the data and the classification task, then using the induced classifier on the labeled set ... |

761 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ... Kumaran, 1996; Williams and Jacobs, 1997), region growing (Hong and Rosenfeld, 1984; Adams and Bischof, 1994), region competition (Zhu and Yuille, 1996), Markov random fields (Geman and Geman, 1984; =-=Blake and Zisserman, 1987-=-), level-set and variational methods (Sethian, 1996), graph cuts (Shi and Malik, 2000), etc. Figure 1.2: Poor imaging conditions and the interactions between foreground and background can confuse a se... |

749 | What Energy Functions can be Minimized via Graph Cuts - Kolmogorov, Zabih - 2004 |

728 | Transductive inference for text classification using support vector machines
- Joachims
- 1999
(Show Context)
Citation Context ...s domain can be solved algebraically with little extra cost. Our work is also closely linked to the transduction problem, the goal of which is to complete the labeling of a partially labeled dataset (=-=Joachims, 1999-=-; Jaakkola et al., 1999; Nigam et al., 1999; Szummer and Jaakkola, 2001). If the labeled data set is rich enough to characterize both the structures of the data and the classification task, then using... |

622 | A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics
- Martin, Fowlkes, et al.
- 2001
(Show Context)
Citation Context ...strained cuts. Using the same spatial mask and the same set of parameters for computing pixel affinity, we apply our constrained normalized cuts to other fashion pictures and Berkeley image datasets (=-=Martin et al., 2001-=-). See Fig 5.10 and Fig 5.11. The number of classes K is chosen manually. When there is an object in the center of the image, such spatial priors always help a segmentation to pick out the object. If ... |

613 |
Feature extraction from faces using deformable templates
- Yuille, Cohen, et al.
- 1992
(Show Context)
Citation Context ... the deformable template approach (Xu et al., 2000), a deformable prototype is used with a deformation space modeled from training data. Some well-known applications are: detecting the eye and mouth (=-=Yuille et al., 1989-=-), tracking shapes in motion (Blake and Isard, 1998), and segmenting anatomical parts in medical images (McInerney and 119sTerzopoulos, 1996). An alternative to deformable templates for object segment... |

567 |
Scale-space filtering
- Witkin
- 1983
(Show Context)
Citation Context ...onsequently, earlier attempts at image segmentation focused on simple low-level feature analysis and enhancement, giving rise to approaches such as thresholding (Sahoo et al., 1988) and edge linking (=-=Witkin, 1983-=-; Canny, 1986; Deriche, 1990). In a basic thresholding approach, a histogram of intensities for all the pixels in the image is first collected. Then a threshold is found by locating the deepest valley... |

478 | Some informational aspects of visual perception
- Attneave
- 1954
(Show Context)
Citation Context ...6), achieving perceptual constancy (Adelson, 1999; Adelson and Pentland, 1996; Hochberg and Brooks, 1962; Lowe, 1984), or compressing the redundant representation by coding only relevant information (=-=Attneave, 1954-=-; Barlow, 1960; Mumford, 1996). Although a sequential system can relieve later stages of perceptual processing of computational burden, such a feed-forward system is vulnerable to mistakes made at eac... |

478 | Deformable models in medical image analysis: A survey
- McInerney, Terzopoulos
- 1996
(Show Context)
Citation Context ...xamples of such segmentation schemes include model-based inference methods: Hough transforms (Illingworth and Kittler, 1988), geometric hashing (Wolfson and Rigoutsos, 1997) and deformable templates (=-=McInerney and Terzopoulos, 1996-=-). The distinction between segmentation and recognition becomes vague. Figure 1.3: Clutter and occlusion pose another major challenge in image segmentation. As a result, significant portions of region... |

475 |
Perceptual Organization and Visual Recognition
- Lowe
- 1985
(Show Context)
Citation Context ...l phenomena (Hoffman, 1983; Witkin and Tenenbaum, 1983; Freeman, 1996; Knill and Richards, 1996), achieving perceptual constancy (Adelson, 1999; Adelson and Pentland, 1996; Hochberg and Brooks, 1962; =-=Lowe, 1984-=-), or compressing the redundant representation by coding only relevant information (Attneave, 1954; Barlow, 1960; Mumford, 1996). Although a sequential system can relieve later stages of perceptual pr... |

445 | Learning to detect natural image boundaries using local brightness, color, and texture cues
- Martin, Fowlkes, et al.
(Show Context)
Citation Context ...y.edu/projects/vision/grouping/segbench/pb/index.html Figure 1.6: Low-level image segmentation alone often does not respect object boundaries. a: Shown here is an image example for edge detection by (=-=Martin et al., 2002-=-). b: edges by Canny edge detector, the goal of which is to detect sharp changes in pixel intensities. It fires strongly in textured areas (e.g. the stony ground), although such intensity changes do n... |

393 |
Seeded region growing
- Adams, Bischof
- 1994
(Show Context)
Citation Context ..., 1988; Cohen, 1991; Ronfard, 1994), boundary detection (Mumford and Shah, 1985; Nitzberg et al., 1993; Geiger and Kumaran, 1996; Williams and Jacobs, 1997), region growing (Hong and Rosenfeld, 1984; =-=Adams and Bischof, 1994-=-), region competition (Zhu and Yuille, 1996), Markov random fields (Geman and Geman, 1984; Blake and Zisserman, 1987), level-set and variational methods (Sethian, 1996), graph cuts (Shi and Malik, 200... |

384 |
A survey of the hough transform
- Illingworth, Kittler
- 1988
(Show Context)
Citation Context ...objects and their relationships to the environment has to be incorporated into the segmentation process. Examples of such segmentation schemes include model-based inference methods: Hough transforms (=-=Illingworth and Kittler, 1988-=-), geometric hashing (Wolfson and Rigoutsos, 1997) and deformable templates (McInerney and Terzopoulos, 1996). The distinction between segmentation and recognition becomes vague. Figure 1.3: Clutter a... |

370 |
Fundamentals of Digital Image processing
- Jain
(Show Context)
Citation Context ...Row #2: edge magnitudes obtained at a fixed scale. Same convention for Fig 1.2 to Fig 1.4. make up for the ignored spatial aspect of segmentation in thresholding approaches, morphological operations (=-=Jain, 1989-=-) are often used subsequently, which poses a tradeoff between removing noise and preserving small structures. Compared to thresholding techniques, edge linking is a spatially local approach, where the... |

354 | Constrained k-means clustering with background knowledge
- Wagstaff, Cardie, et al.
- 2001
(Show Context)
Citation Context ... motivated by the gap between discriminative and generative approaches, we are aware of other works that put similar 116sconstraints into clustering algorithms such as K-means (Wagstaff et al., 2000; =-=Wagstaff et al., 2001-=-). Two types of constraints, must-link and cannot-link, are considered. An earlier version of our work (Yu and Shi, 2001a) also considered cannot-link constraints, that is, two nodes cannot assume the... |

346 |
Exact Maximum A Posteriori Estimation for Binary Images
- Greig, Porteous, et al.
- 1989
(Show Context)
Citation Context ...1). Such pairwise comparisons, however, often have difficulty in deriving reliable long-range grouping information. 115sAttempts have been made to find MRF solutions by graph partitioning algorithms (=-=Greig et al., 1989-=-; Ferrari et al., 1995; Boykov et al., 1998; Roy and Cox, 1998; Ishikawa and Geiger, 1998). In particular, sufficient and necessary conditions on the properties of energy functions that can be solved ... |

338 | Preattentive texture discrimination with early vision mechanisms - Malik, Perona - 1990 |

332 | Segmentation using eigenvectors: a unifying view
- Weiss
- 1999
(Show Context)
Citation Context ...e bipartitioning (Shi and Malik, 2000) or hypercube partitioning (Hendrickson and Leland, 1995). To get K flat partitions, most prior work uses K eigenvectors (Chan et al., 1994; Shi and Malik, 2000; =-=Weiss, 1999; N-=-g 18 £ £ £set al., 2002), others use more eigenvectors than partitions required (Alpert et al., 1995; Malik et al., 2001). Sometimes, the first trivial eigenvector is discarded (Hall, 1970; Shi and... |

317 | Contour and texture analysis for image segmentation
- MALIK, BELONGIE, et al.
- 2001
(Show Context)
Citation Context ...l in applications such as circuit layout (Chan et al., 1994; Alpert and Kahng, 1995a), load balancing in parallel computation (Hendrickson and Leland, 1995) and image segmentation in computer vision (=-=Malik et al., 2001-=-). However, the conceptual simplicity of such approaches is lost in the last step which involves recovering a discrete solution from the continuous solution. A majority of the theoretical work on spec... |

286 | An optimal graph theoretic approach to data clustering: Theory and its application to image segmentation
- Wu, Leahy
- 1993
(Show Context)
Citation Context ...rity, continuity and symmetry, can be first evaluated through a comparison of feature values associated with the elements, then combined into a measure summarizing the overall grouping compatibility (=-=Wu and Leahy, 1993-=-; Shi and Malik, 1997; Puzicha et al., 1998; Gdalyahu et al., 1998; Sharon et al., 2000). While Gestalt laws have always stressed the similarity of elements in grouping, the effect of local feature co... |

285 |
A Suivey of Thresholding Techniques
- Sahoo, Soltani, et al.
(Show Context)
Citation Context ... 1938; Kanizsa, 1979; Palmer, 1999). Consequently, earlier attempts at image segmentation focused on simple low-level feature analysis and enhancement, giving rise to approaches such as thresholding (=-=Sahoo et al., 1988-=-) and edge linking (Witkin, 1983; Canny, 1986; Deriche, 1990). In a basic thresholding approach, a histogram of intensities for all the pixels in the image is first collected. Then a threshold is foun... |

266 | Multi-Camera Scene Reconstruction via Graph Cuts
- Kolmogorov, Zabih
- 2002
(Show Context)
Citation Context ..., 1998; Roy and Cox, 1998; Ishikawa and Geiger, 1998). In particular, sufficient and necessary conditions on the properties of energy functions that can be solved by minimum cuts have been proven in (=-=Kolmogorov and Zabih, 2002-=-b; Ishikawa, 2003). The work here shows that prior knowledge can be used to guide grouping for discriminative criteria such as normalized cuts (Shi and Malik, 1997), and that their global optima in th... |

248 |
Boundary detection by minimizing functionals
- Mumford, Shah
- 1985
(Show Context)
Citation Context ...of such generic priors on boundary and/or region properties is used in various approaches of image segmentation: 2sactive contours (Kass et al., 1988; Cohen, 1991; Ronfard, 1994), boundary detection (=-=Mumford and Shah, 1985-=-; Nitzberg et al., 1993; Geiger and Kumaran, 1996; Williams and Jacobs, 1997), region growing (Hong and Rosenfeld, 1984; Adams and Bischof, 1994), region competition (Zhu and Yuille, 1996), Markov ran... |

240 | A maximum-flow formulation of the n-camera stereo correspondence problem
- Cox
- 1998
(Show Context)
Citation Context ...l grouping. For some particular forms of ε, such as the above mentioned probability criteria using generative models and minimum cuts criteria in discriminative approaches (Ishikawa and Geiger, 1998;=-= Roy and Cox, 1998-=-; Boykov et al., 1999), the constraints in Eqn (5.3) can be trivially incorporated in an algorithm that optimizes the objective. For the former, Markov Chain Monte Carlo is a general solution techniqu... |

231 | Image segmentation by data-driven Markov chain Monte Carlo
- TU, ZHU
- 2002
(Show Context)
Citation Context ...y input with known patterns. It has inspired a whole range of generative approaches. One of the most successful applications of this “analysis = synthesis” idea on segmenting real images is given =-=in (Tu and Zhu, 2002-=-). However, despite its rigorous theoretical basis, the implementation often involves too many parameters and heuristics. In this thesis, we adopt the interactive processing point of view for the abov... |

218 | Partially labeled classification with markov random walks
- Szummer, Jaakkola
- 2002
(Show Context)
Citation Context .... Our work is also closely linked to the transduction problem, the goal of which is to complete the labeling of a partially labeled dataset (Joachims, 1999; Jaakkola et al., 1999; Nigam et al., 1999; =-=Szummer and Jaakkola, 2001-=-). If the labeled data set is rich enough to characterize both the structures of the data and the classification task, then using the induced classifier on the labeled set and interpolating it to the ... |

213 |
Representations of knowledge in complex systems (with discussion
- Grenander, Miller
- 1994
(Show Context)
Citation Context ...ne to false alarms. In contrast, the interactive processing point of view acknowledges the chickenand-egg nature of perceptual organization (Kelly and Grossberg, 2000; Rumelhart and McClelland, 1986; =-=Grenander and Miller, 1994-=-), and overcomes the issue by engaging all perceptual processing simultaneously. The complicating issue that arises is the interactions among the various perceptual modules. Despite agreement on the n... |

205 |
Level Set Methods
- Sethian
- 1996
(Show Context)
Citation Context ...g and Rosenfeld, 1984; Adams and Bischof, 1994), region competition (Zhu and Yuille, 1996), Markov random fields (Geman and Geman, 1984; Blake and Zisserman, 1987), level-set and variational methods (=-=Sethian, 1996-=-), graph cuts (Shi and Malik, 2000), etc. Figure 1.2: Poor imaging conditions and the interactions between foreground and background can confuse a segmentation algorithm based solely on feature simila... |

204 | Filters, random fields and maximum entropy (FRAME) – towards the unified theory for texture modeling
- Zhu, Wu, et al.
- 1996
(Show Context)
Citation Context ...Adelson, 1988; Malik and Perona, 1990). Some other approaches of texture segmentation go beyond the analysis of features obtained from image filters by also modeling the interactions between filters (=-=Zhu et al., 1998-=-). These Markov random field models (Geman and Geman, 1984) capture contextual dependencies and other statistical characteristics of texture features. a: boundary. b: incoherent. c: disconnected. Figu... |

198 |
E.: Neural dynamics of form perception: boundary completion, illusory figures and neon color spreading
- Grossberg, Mingolla
- 1985
(Show Context)
Citation Context ...lated to image segmentation and has been studied from both contour processing and region processing perspectives. Contour-based approaches perform contour completion by using good curve continuation (=-=Grossberg and Mingolla, 1985-=-; Heitger and von der Heydt, 1993; Mumford, 1993; Ullman, 1976; Williams and Jacobs, 1997), whereas region-based approaches perform image partitioning by using surface properties (Nitzberg et al., 199... |

198 | Hierarchical bayesian inference in the visual cortex
- Lee, Mumford
(Show Context)
Citation Context ...y peceptual pheonomena. However, it is overly flexible and lacks a clear formulation of the overall computation. Some theories focus on interpreting the computation carried out in biological systems (=-=Lee and Mumford, 2003-=-) with high-level concepts only. Parallel distributed processing theory (Rumelhart and McClelland, 1986) is again a biologically inspired computational framework for perceptual processing. It has revo... |

188 |
An improved spectral graph partitioning algorithm for mapping parallel computations
- Hendrickson, Leland
- 1995
(Show Context)
Citation Context ...e, numerically efficient (Anstreicher and Wolkowicz, 2000), and successful in applications such as circuit layout (Chan et al., 1994; Alpert and Kahng, 1995a), load balancing in parallel computation (=-=Hendrickson and Leland, 1995-=-) and image segmentation in computer vision (Malik et al., 2001). However, the conceptual simplicity of such approaches is lost in the last step which involves recovering a discrete solution from the ... |

178 | Markov Random Fields with Efficient Approximations
- Boykov, Veksler, et al.
- 1998
(Show Context)
Citation Context ...ten have difficulty in deriving reliable long-range grouping information. 115sAttempts have been made to find MRF solutions by graph partitioning algorithms (Greig et al., 1989; Ferrari et al., 1995; =-=Boykov et al., 1998-=-; Roy and Cox, 1998; Ishikawa and Geiger, 1998). In particular, sufficient and necessary conditions on the properties of energy functions that can be solved by minimum cuts have been proven in (Kolmog... |

175 | Class-specific, top-down segmentation
- Borenstein, Ullman
- 2002
(Show Context)
Citation Context ...motion (Blake and Isard, 1998), and segmenting anatomical parts in medical images (McInerney and 119sTerzopoulos, 1996). An alternative to deformable templates for object segmentation is proposed in (=-=Borenstein and Ullman, 2002-=-). Instead of a globally constrained template, object knowledge is represented using pairs of image fragments and their figureground labeling from a training set. The problem of segmentation becomes o... |

170 |
An R-Dimensional Quadratic Placement Algorithm
- Hall
- 1970
(Show Context)
Citation Context ..., 2000; Weiss, 1999; Ng 18 £ £ £set al., 2002), others use more eigenvectors than partitions required (Alpert et al., 1995; Malik et al., 2001). Sometimes, the first trivial eigenvector is discarde=-=d (Hall, 1970-=-; Shi and Malik, 1997). Some works use eigenvectors literally as point coordinates (Shi and Malik, 2000; Alpert and Kahng, 1995b), while many normalize these points to have unit lengths to construct i... |

164 | C.: Clustering with Instance-level Constraints
- Wagstaff, Cardie
- 2000
(Show Context)
Citation Context ... our work was initially motivated by the gap between discriminative and generative approaches, we are aware of other works that put similar 116sconstraints into clustering algorithms such as K-means (=-=Wagstaff et al., 2000-=-; Wagstaff et al., 2001). Two types of constraints, must-link and cannot-link, are considered. An earlier version of our work (Yu and Shi, 2001a) also considered cannot-link constraints, that is, two ... |

157 | Exact Optimization for Markov Random Fields with Convex Priors
- Ishikawa
- 2003
(Show Context)
Citation Context ...hikawa and Geiger, 1998). In particular, sufficient and necessary conditions on the properties of energy functions that can be solved by minimum cuts have been proven in (Kolmogorov and Zabih, 2002b; =-=Ishikawa, 2003-=-). The work here shows that prior knowledge can be used to guide grouping for discriminative criteria such as normalized cuts (Shi and Malik, 1997), and that their global optima in the continuous doma... |

156 | Geometric hashing: an overview
- Wolfson, Rigoutsos
- 1997
(Show Context)
Citation Context ...has to be incorporated into the segmentation process. Examples of such segmentation schemes include model-based inference methods: Hough transforms (Illingworth and Kittler, 1988), geometric hashing (=-=Wolfson and Rigoutsos, 1997-=-) and deformable templates (McInerney and Terzopoulos, 1996). The distinction between segmentation and recognition becomes vague. Figure 1.3: Clutter and occlusion pose another major challenge in imag... |

154 | A factorization approach to grouping
- Perona, Freeman
- 1998
(Show Context)
Citation Context ...ilarity to pop out. With negative ground-ground connections arising from fragmented background, regularization allows a coherent foreground to stand out. The problem of fragmented background has led (=-=Perona and Freeman, 1998-=-) to adopt an unbalanced criterion which emphasizes the coherence within the figure but not the ground. However, an unbalanced criterion tends to favor small local clusters and thus miss global groupi... |

147 |
Preattentive processing in vision
- TREISMAN
- 1985
(Show Context)
Citation Context ...00). This view of feature discrimination for grouping is supported by evidence in neurophysiology on elaborate feature detectors in visual cortex (van Essen, 1985), in psychophysics on visual search (=-=Treisman, 1985-=-) and in modeling on texture segmentation (Julesz, 1984; Bergen and Adelson, 1988; Malik and Perona, 1990). Some other approaches of texture segmentation go beyond the analysis of features obtained fr... |

146 |
Harmonic Analysis on Semigroups
- Berg, Christensen, et al.
- 1984
(Show Context)
Citation Context ...tinct K columns of V are locally optimal only if the weight matrix is semi-positive definite. Though a pairwise weight matrix evaluated from a Gaussian function is guaranteed to be positive definite (=-=Berg et al., 1984-=-), which is what we will use in our experiments, in general it may not be semi-positive definite. In other words, the objective function may not be convex. However, the global-optimality holds even wh... |

145 | Neuronal responses to static texture patterns in area V1 of the alert macaque monkey - KNIERIM, ESSEN - 1991 |