## A.Blake. Cosegmentation of image pairs by histogram matching - incorporating a global constraint into MRFs (2006)

### Cached

### Download Links

Venue: | In CVPR |

Citations: | 97 - 3 self |

### BibTeX

@INPROCEEDINGS{Rother06a.blake.cosegmentation,

author = {Carsten Rother and Vladimir Kolmogorov and Tom Minka and Andrew Blake},

title = {A.Blake. Cosegmentation of image pairs by histogram matching - incorporating a global constraint into MRFs},

booktitle = {In CVPR},

year = {2006},

pages = {993--1000}

}

### Years of Citing Articles

### OpenURL

### Abstract

We introduce the term cosegmentation which denotes the task of segmenting simultaneously the common parts of an image pair. A generative model for cosegmentation is presented. Inference in the model leads to minimizing an energy with an MRF term encoding spatial coherency and a global constraint which attempts to match the appearance histograms of the common parts. This energy has not been proposed previously and its optimization is challenging and NP-hard. For this problem a novel optimization scheme which we call trust region graph cuts is presented. We demonstrate that this framework has the potential to improve a wide range of research: Object driven image retrieval, video tracking and segmentation, and interactive image editing. The power of the framework lies in its generality, the common part can be a rigid/non-rigid object (or scene), observed from different viewpoints or even similar objects of the same class. 1.

### Citations

2619 | Normalized cuts and image segmentation
- Shi, Malik
- 1997
(Show Context)
Citation Context ...n operation that we term “cosegmentation”. Powerful procedures for low-level segmentation can be produced by incorporating difference measures at the level of pixels, into a global objective function =-=[20, 3, 17]-=-. The objective function can also incorporate a tendency to coherence of regions. Completely automatic segmentation is possible [20] but prone to error, and interactive input [3, 17] or fusion with ot... |

818 | An Experimental Comparison of Min-Cut/Max-Flow Algorithms for Energy Minimization in Vision
- Boykov, Kolmogorov
- 2004
(Show Context)
Citation Context ...ive contours [12, 9]. We argue, however, that since the MRF term is an essential part of the energy, it is desirable to use the well-established technique for binary MRFs - min cut/max flow algorithm =-=[4]-=-. Fortunately, the form of our global term will allow to use max flow algorithm inside the method called submodular-supermodular procedure [15]. For simplicity, in this paper we set c1 = c2 = 1, which... |

749 |
Nonlinear Programming. Athena Scientific
- Bertsekas
- 1995
(Show Context)
Citation Context ...ch are not upper bounds. Furthermore, in this method function E2(x) can also be arbitrary - it is no longer required to be supermodular. Trust region methods are well-known in continuous optimization =-=[2]-=-; TRGC can be viewed as their discrete analogue. A related continuous optimization method is the linearization method of Pschenichnyj [16]. Description of TRGC Instead of selecting unary potentials y ... |

670 | GrabCut": Interactive Foreground Extraction using Iterated Graph Cuts
- Rother, Kolmogorov, et al.
- 2004
(Show Context)
Citation Context ...n operation that we term “cosegmentation”. Powerful procedures for low-level segmentation can be produced by incorporating difference measures at the level of pixels, into a global objective function =-=[20, 3, 17]-=-. The objective function can also incorporate a tendency to coherence of regions. Completely automatic segmentation is possible [20] but prone to error, and interactive input [3, 17] or fusion with ot... |

667 | Interactive Graph Cuts for Optimal Boundary and Region Segmentation of Objects in ND Images
- Boykov, Jolly
- 2001
(Show Context)
Citation Context ...n operation that we term “cosegmentation”. Powerful procedures for low-level segmentation can be produced by incorporating difference measures at the level of pixels, into a global objective function =-=[20, 3, 17]-=-. The objective function can also incorporate a tendency to coherence of regions. Completely automatic segmentation is possible [20] but prone to error, and interactive input [3, 17] or fusion with ot... |

607 | Relations between the statistics of natural images and the response properties of cortical cells
- Field
- 1987
(Show Context)
Citation Context ...annon divergence between foreground histograms (see [18]). However, the independence assumption is something of a drawback, as it is known that nearby pixels in an image are not generally independent =-=[8]-=-. If instead we choose a generative model for the foreground histograms as a whole, rather than individual pixels, we can obtain other standard divergences such as variational distance. A further Isin... |

564 | Kernel-based object tracking
- Comaniciu, Ramesh, et al.
(Show Context)
Citation Context ...ises in cosegmentation is the problem of finding a coherent image region with given target histogram. This problem has been approached previously using ellipses or active contours to define coherence =-=[6, 12, 9]-=-. Inspired by [17], we instead define coherence via MRF priors and solve the problem with iterated graph cuts. In order to arrive at an objective function for cosegmentation, we begin, in sec. 2, by s... |

399 | SIMPLIcity: Semantics-sensitive Integrated Matching for Picture LIbraries
- Wang, Li, et al.
- 2001
(Show Context)
Citation Context ...mages has been computed by comparing the global statistics of the two images. Typically the comparison is applied to the histograms of each image, constructed from features such as colour and texture =-=[11, 7]-=-. However, such a global, undifferentiated approach to comparison is liable to result in crude comparisons, as figures 5, 6 show. Apparently, it is essential to incorporate some form of differentiatio... |

159 | Locus: Learning object classes with unsupervised segmentation
- Winn, Jojic
- 2005
(Show Context)
Citation Context ... other modalities [13], is normally needed to correct those errors. Another source of information for correcting segmentation is to supply a database of related images and segment them simultaneously =-=[21]-=-. Here we demonstrate that supplying just one additional image can be sufficient to segment both together, to higher accuracy than is achieved with either one alone. Furthermore, in contrast to [21] w... |

158 | Mean Shift Analysis and Applications
- Comaniciu, Meer
- 1999
(Show Context)
Citation Context ...of images could be captured despite the subjects being unrelated.) One approach to such differentiation, is “integrated region matching” [11], in which images are subjected to mean-shift segmentation =-=[5]-=-, and then a simple similarity measure records the similarity of paired regions, in a search over both segmented images. However, the choice of paired regions takes no account of object coherence, and... |

137 | A combinatorial algorithm minimizing submodular functions in strongly polynomial time
- Schrijver
- 2000
(Show Context)
Citation Context ...x ′ ) + E1(x ∨ x ′ ) ≤ E1(x) + E1(x ′ ) E2(x ∧ x ′ ) + E2(x ∨ x ′ ) ≥ E2(x) + E2(x ′ ) for all configurations x,x ′ . It is well-known that any submodular function can be minimized in polynomial time =-=[19]-=-. In our case E1(x) is a sum of unary and pairwise terms, so a global minimum of E1 can be computed very efficiently via min cut/max flow algorithm. The presence of supermodular part, however, makes t... |

103 | Recognizing surfaces using three-dimensional textons
- Leung, Malik
- 1999
(Show Context)
Citation Context ... 0.3, λ1 = 1, λ2 = 50 and b = 0.5. Finally let us introduce our appearance model. We have experimented with a simple 2D intensity normalized RGB colour space and a richer texture (texton) based model =-=[14]-=-, which has been proven to be very powerful for image retrieval [7]. Apart from scenarios of retrieving images of the same class we have used the simple model since the emphases on colour improved the... |

56 | Bilayer segmentation of binocular stereo video
- Kolmogorov, Criminisi, et al.
- 2005
(Show Context)
Citation Context ... function can also incorporate a tendency to coherence of regions. Completely automatic segmentation is possible [20] but prone to error, and interactive input [3, 17] or fusion with other modalities =-=[13]-=-, is normally needed to correct those errors. Another source of information for correcting segmentation is to supply a database of related images and segment them simultaneously [21]. Here we demonstr... |

53 | Generalizing swendsen-wang to sampling arbitrary posterior probabilities
- Barbu, Zhu
- 2007
(Show Context)
Citation Context ...this global term makes the minimization problem quite challenging. One could think of using some general inference algorithm, such as Swendsen-Wang Cuts for sampling arbitrary posterior probabilities =-=[1]-=-. Anotherspossibility is to use active contours [12, 9]. We argue, however, that since the MRF term is an essential part of the energy, it is desirable to use the well-established technique for binary... |

45 | A.: The concave-convex procedure (CCCP
- Yuille, Rangarajan
- 2001
(Show Context)
Citation Context ...ses an alternative method - trust region graph cuts. 3.1. Submodular-supermodular procedure (SSP) This method was inspired by concave-convex procedure for minimizing functions of continuous variables =-=[22]-=-. SSP is an iterative technique which produces a sequence of configurations x 0 ,x 1 , . . .,x t , . . .. The main property of SSP is that the energy never goes up, i.e. E(x 0 ) ≥ E(x 1 ) ≥ . . .. Let... |

39 | Active contours for tracking distributions
- Freedman, Zhang
- 2004
(Show Context)
Citation Context ...ises in cosegmentation is the problem of finding a coherent image region with given target histogram. This problem has been approached previously using ellipses or active contours to define coherence =-=[6, 12, 9]-=-. Inspired by [17], we instead define coherence via MRF priors and solve the problem with iterated graph cuts. In order to arrive at an objective function for cosegmentation, we begin, in sec. 2, by s... |

16 | Classification error rate for quantitative evaluation of content-based image retrieval systems
- Deselaers, Keysers, et al.
(Show Context)
Citation Context ...mages has been computed by comparing the global statistics of the two images. Typically the comparison is applied to the histograms of each image, constructed from features such as colour and texture =-=[11, 7]-=-. However, such a global, undifferentiated approach to comparison is liable to result in crude comparisons, as figures 5, 6 show. Apparently, it is essential to incorporate some form of differentiatio... |

16 | Shape gradients for histogram segmentation using active contours
- Jehan-Besson, Barlaud, et al.
(Show Context)
Citation Context ...ises in cosegmentation is the problem of finding a coherent image region with given target histogram. This problem has been approached previously using ellipses or active contours to define coherence =-=[6, 12, 9]-=-. Inspired by [17], we instead define coherence via MRF priors and solve the problem with iterated graph cuts. In order to arrive at an objective function for cosegmentation, we begin, in sec. 2, by s... |

13 |
A.H.: Exact MAP estimation for binary images
- Greig, Porteous, et al.
- 1989
(Show Context)
Citation Context ...h boundaries. The optimisation of the objective function arising from that generative model, is something of a challenge. Graph cut algorithms are widely used for binary optimisation in Markov models =-=[10, 3]-=-, but have not been used before where the objective function contains a histogram difference measure. It transpires that such an objective function is not “submodular” and therefore strictly not tract... |

8 | The Linearization Method for Constrained Optimization - Pshenichnyj - 1994 |

6 |
A Supermodular-Submodular Procedure with Applications to Discriminative Structure Learning
- Narasimhan, Bilmes
- 2005
(Show Context)
Citation Context ...ed technique for binary MRFs - min cut/max flow algorithm [4]. Fortunately, the form of our global term will allow to use max flow algorithm inside the method called submodular-supermodular procedure =-=[15]-=-. For simplicity, in this paper we set c1 = c2 = 1, which means that we prefer foreground regions of the same size. It is easy, however, to extend the model to account for different sizes: we can put ... |