## Mean Shift Based Clustering in High Dimensions: A Texture Classification Example (2003)

### Cached

### Download Links

Citations: | 100 - 2 self |

### BibTeX

@MISC{Georgescu03meanshift,

author = {Bogdan Georgescu and Ilan Shimshoni and Peter Meer},

title = {Mean Shift Based Clustering in High Dimensions: A Texture Classification Example},

year = {2003}

}

### Years of Citing Articles

### OpenURL

### Abstract

Feature space analysis is the main module in many computer vision tasks. The most popular technique, k-means clustering, however, has two inherent limitations: the clusters are constrained to be spherically symmetric and their number has to be known a priori. In nonparametric clustering methods, like the one based on mean shift, these limitations are eliminated but the amount of computation becomes prohibitively large as the dimension of the space increases. We exploit a recently proposed approximation technique, locality-sensitive hashing (LSH), to reduce the computational complexity of adaptive mean shift. In our implementation of LSH the optimal parameters of the data structure are determined by a pilot learning procedure, and the partitions are data driven. As an application, the performance of mode and k-means based textons are compared in a texture classification study.

### Citations

2630 | Density Estimation for Statistics and Data Analysis (Chapman and - Silverman - 1986 |

1618 | Mean shift: A robust approach toward feature space analysis
- Comaniciu, Meer
(Show Context)
Citation Context ...s the modes of the multivariate distribution underlying the feature space. The number of clusters is obtained automatically by finding the centers of the densest regions in the space (the modes). See =-=[1]-=- for details. Under its original implementation the mean shift based clustering cannot be used in high dimensional spaces. Already for � ��, in a video sequence segmentation application, a fineto-coar... |

1044 | Computational Geometry: Algorithms and Applications - Berg, Kreveld, et al. - 1997 |

762 | Approximate nearest neighbors: Towards removing the curse of dimensionality
- Indyk, Motwani
- 1998
(Show Context)
Citation Context .... Recently, several probabilistic algorithms have been proposed for approximate nearest neighbor search. The algorithms yield sublinear complexity with a speedup which depends on the desired accuracy =-=[7, 10, 11]-=-. In this paper we have adapted the algorithm in [7] for mean shift based clustering in high dimensions. Working with data in high dimensions also required that we extend the adaptive mean shift proce... |

461 | Similarity search in high dimensions via hashing
- Gionis, Indyk, et al.
- 1999
(Show Context)
Citation Context .... Recently, several probabilistic algorithms have been proposed for approximate nearest neighbor search. The algorithms yield sublinear complexity with a speedup which depends on the desired accuracy =-=[7, 10, 11]-=-. In this paper we have adapted the algorithm in [7] for mean shift based clustering in high dimensions. Working with data in high dimensions also required that we extend the adaptive mean shift proce... |

352 |
Representing and recognizing the visual appearance of materials using three-dimensional textons
- Leung, Malik
(Show Context)
Citation Context ...o (h). The parameters: MS – bandwidth �; AMS – number of neighbors �. 5. Texture Classification Efficient methods exist for texture classification under varying illumination and viewing direction [3],=-=[12]-=-, [15], [18]. In the state-of-the-art approaches a texture is characterized through textons, which are cluster centers in a feature space derived from the input. Following [12] this feature space is b... |

202 | Fast pose estimation with parameter-sensitive hashing
- Shakhnarovich, Viola, et al.
- 2003
(Show Context)
Citation Context ...lity-sensitive hashing (LSH) [7] and adapted it to handle the complex data met in computer vision applications. In a task of estimating the pose of articulated objects, described in these proceedings =-=[16]-=-, the LSH technique was extended to accommodate distances in the parameter space. 3.1. High Dimensional Neighborhood Queries Given Ò points in Ê� the mean shift iterations (5) require a neighborhood q... |

196 | Efficient search for approximate nearest neighbor in high dimensional spaces
- Kushilevitz, Ostrovsky, et al.
(Show Context)
Citation Context .... Recently, several probabilistic algorithms have been proposed for approximate nearest neighbor search. The algorithms yield sublinear complexity with a speedup which depends on the desired accuracy =-=[7, 10, 11]-=-. In this paper we have adapted the algorithm in [7] for mean shift based clustering in high dimensions. Working with data in high dimensions also required that we extend the adaptive mean shift proce... |

135 | A simple algorithm for nearest neighbor search in high dimensions
- Nene, Nayar
- 1997
(Show Context)
Citation Context ... need for a fast algorithm to perform neighborhood queries when computing (5). The problem has been addressed before in the vision community by sorting the data according to each of the � coordinates =-=[13]-=-, but a significant speedup was achieved only when the data is close to a low-dimensional manifold. Recently new algorithms using tools from probabilistic approximation theory were suggested for perfo... |

133 | Texture classification: are filter banks necessary
- Varma, Zisserman
- 2003
(Show Context)
Citation Context ...which are cluster centers in a feature space derived from the input. Following [12] this feature space is built from the output of a filter bank applied at every pixel. However, as was shown recently =-=[19]-=-, neighborhood information in the spatial domain may also suffice. The approaches differ in the employed filter bank. – LM: A combination of 48 anisotropic and isotropic filters were used by Leung and... |

111 | The variable bandwidth mean shift and data-driven scale seclection
- Comaniciu, Ramesh, et al.
- 2001
(Show Context)
Citation Context ...have adapted the algorithm in [7] for mean shift based clustering in high dimensions. Working with data in high dimensions also required that we extend the adaptive mean shift procedure introduced in =-=[2]-=-. All computer vision applications of mean shift until now, such as image segmentation, object recognition and tracking, were in relatively low-dimensional spaces. Our implementation opens the door to... |

104 | Constructing models for content-based image retrieval
- Schmid
(Show Context)
Citation Context ... The parameters: MS – bandwidth �; AMS – number of neighbors �. 5. Texture Classification Efficient methods exist for texture classification under varying illumination and viewing direction [3],[12], =-=[15]-=-, [18]. In the state-of-the-art approaches a texture is characterized through textons, which are cluster centers in a feature space derived from the input. Following [12] this feature space is built f... |

71 | Spatial-Temporal Segmentation of Video by Hierarchical Mean Shift Analysis
- DeMenthon, Megret
- 2002
(Show Context)
Citation Context ...n the mean shift based clustering cannot be used in high dimensional spaces. Already for � ��, in a video sequence segmentation application, a fineto-coarse hierarchical approach had to be introduced =-=[5]-=-. The most expensive operation of the mean shift method is finding the closest neighbors of a point in the space. The problem is known in computational geometry as multidimensional range searching [4,... |

64 |
Compact representation of bidirectional texture functions
- Cula, Dana
(Show Context)
Citation Context ...e) to (h). The parameters: MS – bandwidth �; AMS – number of neighbors �. 5. Texture Classification Efficient methods exist for texture classification under varying illumination and viewing direction =-=[3]-=-,[12], [15], [18]. In the state-of-the-art approaches a texture is characterized through textons, which are cluster centers in a feature space derived from the input. Following [12] this feature space... |

50 | Rotation-invariant texture classification using a complete space-frequency model
- Haley, Manjunath
- 1999
(Show Context)
Citation Context ...tion yielded 896 models and 896 queries. The recognition rate decreased for all the filter banks. The best result of 94%, was again obtained with the LM filters for both the mean and mode textons. In =-=[8]-=-, with the same setup but employing a different texture representation, and using only 109 textures from the Brodatz database the recognition rate was � �� . A texture class is characterized by the hi... |

39 | A sublinear time approximation scheme for clustering in metric spaces
- Indyk
- 1999
(Show Context)
Citation Context ...gorithms using tools from probabilistic approximation theory were suggested for performing approximate nearest neighbor search in high dimensions for general datasets [10, 11] and for clustering data =-=[9, 14]-=-. We use the approximate nearest neighbor algorithm based on locality-sensitive hashing (LSH) [7] and adapted it to handle the complex data met in computer vision applications. In a task of estimating... |

32 | and Y.Rabani. Polynomial time approximation schemes for geometric clustering problems
- Ostrovsky
- 2002
(Show Context)
Citation Context ...gorithms using tools from probabilistic approximation theory were suggested for performing approximate nearest neighbor search in high dimensions for general datasets [10, 11] and for clustering data =-=[9, 14]-=-. We use the approximate nearest neighbor algorithm based on locality-sensitive hashing (LSH) [7] and adapted it to handle the complex data met in computer vision applications. In a task of estimating... |

6 |
Classifying images of materials
- Varma, Zisserman
(Show Context)
Citation Context ...arameters: MS – bandwidth �; AMS – number of neighbors �. 5. Texture Classification Efficient methods exist for texture classification under varying illumination and viewing direction [3],[12], [15], =-=[18]-=-. In the state-of-the-art approaches a texture is characterized through textons, which are cluster centers in a feature space derived from the input. Following [12] this feature space is built from th... |