#### DMCA

## Multi-scale EM-ICP: A Fast and Robust Approach for Surface Registration (2002)

### Cached

### Download Links

- [ftp-sop.inria.fr]
- [www.inria.fr]
- [www-sop.inria.fr]
- DBLP

### Other Repositories/Bibliography

Venue: | European Conference on Computer Vision (ECCV 2002), volume 2353 of LNCS |

Citations: | 106 - 7 self |

### Citations

3065 |
A method for registration of 3D shapes
- Besl, McKay
- 1992
(Show Context)
Citation Context ...s paper describes a new method for the registration of surfaces. This kind of registration is usually performed using one of the multiple variations around the Iterative Closest Point (ICP) algorithm =-=[1,19]. Th-=-is algorithm is quite fast and A. Heyden et al. (Eds.): ECCV 2002, LNCS 2353, pp. 418–432, 2002. c○ Springer-Verlag Berlin Heidelberg 2002sMulti-scale EM-ICP: A Fast and Robust Approach 419 accura... |

980 | A view of the EM algorithm that justifies incremental, sparse, and other variants
- Neal, Hinton
- 1999
(Show Context)
Citation Context ...e EM Algorithm In fact, it turns out that the algorithm justified in Sec. 2.3 is no more than the optimisation of the criterion presented in Sec. 2.2, through an ExpectationMaximisation (EM) approach =-=[3,13]. -=-Thus, the algorithm is ensured to converge. To relate the criterion CEM(T )=−log(p(S|M,T)) with the optimisation method, let us introduce the matching matrix in the criterion using Bayes rule and ta... |

613 | Scale-space theory in computer vision
- Lindeberg
- 1993
(Show Context)
Citation Context ...on, and linked to the multi-grid approach with intensity images. Our discussion will be quite pragmatic, though we think it could be integrated into a rigorous multi-scale framework (see for instance =-=[10]-=-).sMulti-scale EM-ICP: A Fast and Robust Approach 427 Coarse-to-fine approaches. The main idea underlying multiscale intensity(i.e. image-) based registration approaches is to get rid of small details... |

283 |
Statistical Optimization for Geometric Computation: Theory and Practice
- Kanatani
- 1996
(Show Context)
Citation Context ...si,mj)+Cte (3) One recognises here the standard ICP criterion using the Mahalanobis distance. This proves that ICP is no more than a maximum likelihood approach of the registration problem. Moreover, =-=[9]-=- showed that this is the best (minimal variance) estimator. Note that the criterion is invariant w.r.t a global scaling of the noise covariance. This property will not hold any more for the following ... |

225 | A pyramid approach to subpixel registration based on intensity
- Thevenaz, Unser
- 1998
(Show Context)
Citation Context ...oaches is to get rid of small details that create local minima and trap the algorithm when the initialization is too far away from the solution. This is especially the case for non-rigid registration =-=[17]-=-. This is usually implemented using a pyramid of blurred images with an increasing Gaussian kernel [8]. Using a Gaussian mixture model for representing the probability of measuring points follows esse... |

150 |
Estimating 3D rigid body transformations: a comparison of four major algorithms
- Eggert, Lorusso, et al.
- 1997
(Show Context)
Citation Context ...e with the ICP for optimizing the transformation is the presence of non-binary weights. In the rigid case, this can be solved by a straightforward adaptation of the SVD or the unit quaternion methods =-=[4]. One can -=-even simplify Eq. (12) by considering that each scene point si contribute to the criterion through � j (AT )ij�T ⋆si − mj� 2 . As the weights are ijs424 S. Granger and X. Pennec normalized, ... |

91 |
A Robust Method for Registration and Segmentation of Multiple Range Images
- Masuda, Yokoya
- 1994
(Show Context)
Citation Context ... obtained through an annealing scheme on this variance parameter. Other works suggest that improvements in terms of speed could be obtained with algorithmic tricks or by decimating the sets of points =-=[12]-=-. Up to our knowledge, none of these works provide guarantees on the convergence of the algorithm nor on the conservation of the robustness and accuracy properties. Our work was inspired by Rangarajan... |

71 | Statistical approaches to feature-based object recognition - Wells - 1997 |

51 |
A Pyramid Framework for Early Vision
- Jolion, Rosenfeld
- 1994
(Show Context)
Citation Context ...zation is too far away from the solution. This is especially the case for non-rigid registration [17]. This is usually implemented using a pyramid of blurred images with an increasing Gaussian kernel =-=[8]-=-. Using a Gaussian mixture model for representing the probability of measuring points follows essentially the same idea: we are representing the pdf of measuring one point of the model as a blurred ve... |

49 |
Registration and integration of multiple range images for 3-d model construction
- Masuda, Sakaue, et al.
- 1996
(Show Context)
Citation Context ...he robustness and the computation times without loosing the accuracy. We found in the literature three main classes of methods to improve the robustness of ICP. The first class uses robust estimators =-=[19,11]-=- to deal with outliers (i.e. erroneous or occulted points). However, these techniques are not designed to be robust w.r.t. the initial transformation, which is our main point of interest in this paper... |

48 | A robust point matching algorithm for autoradiograph alignment
- Rangarajan, Mjolsness, et al.
- 1997
(Show Context)
Citation Context ...esented by Rangarajan et al., who introduced multiple weighted matches justified by a probabilistic vision of the matching problem. They developed matching models based on Gaussian weight (SoftAssign =-=[16]-=-) and Mutual Information [15], leading to a smaller number of local minima and thus presenting the most convincing improvements. The last two classes are usually based on a variance parameter. For hig... |

35 | A feature registration framework using mixture models
- Chui, Rangarajan
- 2000
(Show Context)
Citation Context ...s roughly corresponds to an ICP with multiple matches weighted by normalized Gaussian weights, giving birth to the EM-ICP acronym of the method. A similar approach has been independently developed in =-=[2]-=-. Section 3 investigates the influence of a new parameter: the variance of the Gaussian. We show that EM-ICP robustly aligns the barycenters and inertia moments with a high variance, while it tends to... |

33 | Rigid point feature registration using mutual information
- Rangarajan, Chui, et al.
- 1999
(Show Context)
Citation Context ... who introduced multiple weighted matches justified by a probabilistic vision of the matching problem. They developed matching models based on Gaussian weight (SoftAssign [16]) and Mutual Information =-=[15]-=-, leading to a smaller number of local minima and thus presenting the most convincing improvements. The last two classes are usually based on a variance parameter. For high values, the algorithm is mo... |

11 |
A stochastic iterative closest point algorithm (stochasticp
- Penney, Edwards, et al.
- 2001
(Show Context)
Citation Context ...ints). However, these techniques are not designed to be robust w.r.t. the initial transformation, which is our main point of interest in this paper. The second class is based on stochastic approaches =-=[14]-=-. This kind of variants are quite efficient, but usually require more computation time. The last class is based on the smoothing of the criterion. The most interesting work has been presented by Ranga... |

10 | The EM Algorithm: A Guided Tour
- Couvreur
- 1996
(Show Context)
Citation Context ...e EM Algorithm In fact, it turns out that the algorithm justified in Sec. 2.3 is no more than the optimisation of the criterion presented in Sec. 2.2, through an ExpectationMaximisation (EM) approach =-=[3,13]. -=-Thus, the algorithm is ensured to converge. To relate the criterion CEM(T )=−log(p(S|M,T)) with the optimisation method, let us introduce the matching matrix in the criterion using Bayes rule and ta... |

4 | Rigid point-surface registration using oriented points and an EM variant of ICP for computer guided oral implantology
- Granger, Pennec, et al.
(Show Context)
Citation Context ...imentally observed that the decrease was too sharp and was sticking the algorithm in local minima. 3.2 Criterion Shape and Annealing Scheme To understand the influence of the variance, we analyzed in =-=[7]-=- the asymptotic values of the criterion and showed that, for high values of the variance, it simply aligns inertia centers and moments of the two point sets, while it reaches the standard ICP criterio... |

3 |
Iterative Point Matching for Registration of Free-Form Surfaces
- Zhang
- 1994
(Show Context)
Citation Context ...s paper describes a new method for the registration of surfaces. This kind of registration is usually performed using one of the multiple variations around the Iterative Closest Point (ICP) algorithm =-=[1,19]. Th-=-is algorithm is quite fast and A. Heyden et al. (Eds.): ECCV 2002, LNCS 2353, pp. 418–432, 2002. c○ Springer-Verlag Berlin Heidelberg 2002sMulti-scale EM-ICP: A Fast and Robust Approach 419 accura... |

1 |
et al. A new approach for dental implant aided surgery. a pilot evaluation
- Etienne
- 2000
(Show Context)
Citation Context ... two large sets of points. As our registration method is developed as a component of a per-operative system for surgery guidance namely the DentalNavigator system (patent pending) developed by AREALL =-=[5]-=-, and based on finely sampled surfaces (containing 30000 to 200000 points), we need to drastically improve both the robustness and the computation times without loosing the accuracy. We found in the l... |