## Top Down Image Segmentation using Congealing and Graph-Cut

### BibTeX

@MISC{Moore_topdown,

author = {Douglas Moore and John Stevens and Scott Lundberg and Bruce A. Draper},

title = {Top Down Image Segmentation using Congealing and Graph-Cut},

year = {}

}

### OpenURL

### Abstract

This paper develops a weakly supervised algorithm that learns to segment rigid multi-colored objects from a set of training images and key points. The approach uses congealing to learn a probabilistic spatial model of the multi-colored object class and graph-cut to separate the foreground from the background. The result is a novel approach which can segment heterogeneous objects, in contrast to other recent approaches which are better at segmenting uniform but possibly flexible objects. 1

### Citations

8919 | Maximum Likelihood from Incomplete Data via the EM-Alogrithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ...high entropy). More specifically, the spatial model is created by aligning the source images through congealing [12], and then fitting a two-class (foreground/background) Gaussian Mixture Model (GMM; =-=[2, 4]-=-) to the resulting entropy values. The resulting probabilistic spatial model is used as the predictive component in the graph-cut algorithm [3], which segments images by balancing a-priori expectation... |

5587 | Distinctive image features from scale-invariant keypoints
- Lowe
- 2004
(Show Context)
Citation Context ... seen from similar viewpoints; i.e. the appearance of the object undergoes only affine transformations among images. It also assumes a single ontarget key point per image of the type produced by SIFT =-=[10]-=-, Scale-saliency [8], or any of several interest point operators [11]. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the ou... |

913 | Object class recognition by unsupervised scale-invariant learning
- Fergus, Perona, et al.
- 2003
(Show Context)
Citation Context ...s [11]. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the output of many recent attention-based object recognition systems =-=[5, 6, 7, 19]-=-. This algorithm can therefore be used to segment objects which have been recognized based on selective attention windows. Given a set of images and key points, our algorithm builds a probabilistic sp... |

745 | What energy functions can be minimized via graph cuts
- Kolmogorov, Zabih
(Show Context)
Citation Context ...ition to congealing error, and therefore indirectly to segmentation error. 2 Related Work Research into top-down image segmentation has been revolutionized by the introduction of graph-cut algorithms =-=[9]-=-. Graph-cut provides a well-motivated and computationally efficient method for segmenting images based on (1) a probabilistic predictive model and (2) local edge data. In essence, it finds the segment... |

703 | GrabCut” - interactive foreground extraction using iterated graph cuts
- Rother, Kolmogorov, et al.
- 2004
(Show Context)
Citation Context ...edicted model and the available boundary information. Graph-cut in turn spurred research, including this work, into methods of learning predictive object models for use in top-down image segmentation =-=[13, 17, 18]-=-. Among previous research efforts, the closest to this work is LOCUS [17]. LOCUS begins with a class of unsegmented images which it segments through inference and learning. While able to segment defor... |

697 | Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images
- Boykov, Jolly
- 2001
(Show Context)
Citation Context ... (foreground/background) Gaussian Mixture Model (GMM; [2, 4]) to the resulting entropy values. The resulting probabilistic spatial model is used as the predictive component in the graph-cut algorithm =-=[3]-=-, which segments images by balancing a-priori expectations with image-specific edge information. As shown experimentally in section 4, the algorithm is able to segment objects when the key points are ... |

511 | A Gentle Tutorial of the EM Algorithm and its Applications to Parameter Estimation for Gaussian Mixtures and Hidden Markov Models
- Bilmes
- 1998
(Show Context)
Citation Context ...high entropy). More specifically, the spatial model is created by aligning the source images through congealing [12], and then fitting a two-class (foreground/background) Gaussian Mixture Model (GMM; =-=[2, 4]-=-) to the resulting entropy values. The resulting probabilistic spatial model is used as the predictive component in the graph-cut algorithm [3], which segments images by balancing a-priori expectation... |

496 | Learning generative visual models from few training examples: an incremental bayesian approach tested on 101 object categories
- Fei-Fei, Fergus, et al.
- 2004
(Show Context)
Citation Context ...s [11]. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the output of many recent attention-based object recognition systems =-=[5, 6, 7, 19]-=-. This algorithm can therefore be used to segment objects which have been recognized based on selective attention windows. Given a set of images and key points, our algorithm builds a probabilistic sp... |

248 | A comparison of affine region detectors
- Mikolajczyk
- 2005
(Show Context)
Citation Context ...rgoes only affine transformations among images. It also assumes a single ontarget key point per image of the type produced by SIFT [10], Scale-saliency [8], or any of several interest point operators =-=[11]-=-. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the output of many recent attention-based object recognition systems [5, 6,... |

226 | SVM-KNN: Discriminative nearest neighbor classification for visual category recognition
- Zhang, Berg, et al.
- 2006
(Show Context)
Citation Context ...s [11]. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the output of many recent attention-based object recognition systems =-=[5, 6, 7, 19]-=-. This algorithm can therefore be used to segment objects which have been recognized based on selective attention windows. Given a set of images and key points, our algorithm builds a probabilistic sp... |

162 | LOCUS: Learning object classes with unsupervised segmentation
- Winn, Jojic
- 2005
(Show Context)
Citation Context ...edicted model and the available boundary information. Graph-cut in turn spurred research, including this work, into methods of learning predictive object models for use in top-down image segmentation =-=[13, 17, 18]-=-. Among previous research efforts, the closest to this work is LOCUS [17]. LOCUS begins with a class of unsegmented images which it segments through inference and learning. While able to segment defor... |

97 |
saliency and image description
- Scale
(Show Context)
Citation Context ...ewpoints; i.e. the appearance of the object undergoes only affine transformations among images. It also assumes a single ontarget key point per image of the type produced by SIFT [10], Scale-saliency =-=[8]-=-, or any of several interest point operators [11]. Unlike related approaches, we allow the object to be multi-colored with arbitrary surface markings. These conditions match the output of many recent ... |

97 | Learning from one example through shared densities of transforms
- Miller, Matsakis, et al.
- 2000
(Show Context)
Citation Context ... consistent (low entropy) while background pixels should be comparatively inconsistent (high entropy). More specifically, the spatial model is created by aligning the source images through congealing =-=[12]-=-, and then fitting a two-class (foreground/background) Gaussian Mixture Model (GMM; [2, 4]) to the resulting entropy values. The resulting probabilistic spatial model is used as the predictive compone... |

68 |
A test for normality based on sample entropy
- Vasicek
- 1976
(Show Context)
Citation Context ...from the source images. To quantify how well windows are aligned, a window stack Ω=W (T ) is defined, and the overall entropy of each pixel column Ωx,y is computed using Vasick’s entropy approximator =-=[16]-=- to create an entropy image e. The total entropy of the window stack is the sum of the entropy of the pixel columns Ωx,y as in Equation 1 (P is the set of all valid pixel locations in a window). H(Ω) ... |

48 | Colour Image Segmentation: A Survey
- Skarbek, Koschan
- 1994
(Show Context)
Citation Context ...ge presentations. Object segmentation is the process of labeling image pixels as either object (foreground) or non-object (background), and has been a difficult problem in computer vision for decades =-=[15]-=-. Segmentation techniques fall into two broad categories: bottom up approaches, which only use data from a single image, and top down approaches, which incorporate external information such as categor... |

42 | Concurrent object recognition and segmentation by graph partitioning
- Yu, Gross, et al.
- 2002
(Show Context)
Citation Context ...edicted model and the available boundary information. Graph-cut in turn spurred research, including this work, into methods of learning predictive object models for use in top-down image segmentation =-=[13, 17, 18]-=-. Among previous research efforts, the closest to this work is LOCUS [17]. LOCUS begins with a class of unsegmented images which it segments through inference and learning. While able to segment defor... |

11 |
Implementing the expert object recognition pathway
- Draper, Baek, et al.
- 2003
(Show Context)
Citation Context |

1 |
A Probablistic Model for Object Recognition
- Simon, Seitz
- 2007
(Show Context)
Citation Context ...learning. While able to segment deformable objects, it suffers from one of the same limitations as bottom-up techniques: it assumes that the target objects are internally homogeneous. Simon and Seitz =-=[14]-=- generate a predictive model based on extracted attention 978-1-4244-2175-6/08/$25.00 ©2008 IEEEwindows from a pre-segmented template image. Our approach differs in that our algorithm generates a pre... |