## MLESAC: A New Robust Estimator with Application to Estimating Image Geometry (2000)

### Cached

### Download Links

- [www.cs.cmu.edu]
- [www-2.cs.cmu.edu]
- [imogen.robots.ox.ac.uk:20000]
- DBLP

### Other Repositories/Bibliography

Venue: | Computer Vision and Image Understanding |

Citations: | 241 - 8 self |

### BibTeX

@ARTICLE{Torr00mlesac:a,

author = {P. H. S. Torr and A. Zisserman},

title = {MLESAC: A New Robust Estimator with Application to Estimating Image Geometry},

journal = {Computer Vision and Image Understanding},

year = {2000},

volume = {78},

pages = {2000}

}

### Years of Citing Articles

### OpenURL

### Abstract

A new method is presented for robustly estimating multiple view relations from point correspondences. The method comprises two parts. The first is a new robust estimator MLESAC which is a generalization of the RANSAC estimator. It adopts the same sampling strategy as RANSAC to generate putative solutions, but chooses the solution that maximizes the likelihood rather than just the number of inliers. The second part of the algorithm is a general purpose method for automatically parameterizing these relations, using the output of MLESAC. A difficulty with multiview image relations is that there are often nonlinear constraints between the parameters, making optimization a difficult task. The parameterization method overcomes the difficulty of nonlinear constraints and conducts a constrained optimization. The method is general and its use is illustrated for the estimation of fundamental matrices, image–image homographies, and quadratic transformations. Results are given for both synthetic and real images. It is demonstrated that the method gives results equal or superior to those of previous approaches. c ○ 2000 Academic Press 1.

### Citations

8142 | Maximum likelihood from incomplete data via the EM algorithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ... + ˆy¯ − y i i i j=1,2 (2σ 2 ) + (1 − γ ) 1 v �� . (10) Given a suitable initial estimate there are several ways to estimate the parameters of the mixture model, most prominent being the EM algorithm =-=[2, 20]-=-, but gradient descent methods could also be used. Because of the presence of outliers in the data the standard method of least-squares estimation is often not suitable as an initial estimate, and it ... |

2467 |
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
- FISCHLER, BOLLES
- 1981
(Show Context)
Citation Context ... for the fundamental matrix. It will be seen that any parameterization must enforce this constraint to accurately capture the two-view geometry. Due to the frequent occurrence of mismatches, a RANSAC =-=[4]-=--like robust estimator is used to estimate the two-view relation. The RANSAC algorithm is a hypothesis and verify algorithm. It proceeds by repeatedly generating solutions estimated from minimal sets ... |

1732 | A combined corner and edge detector
- Harris, Stephens
- 1988
(Show Context)
Citation Context ...m image correspondences over two views. This computation requires initial matching of points (corners) over the image pairs. Corners are detected to subpixel accuracy using the Harris corner detector =-=[9]-=-. Given a corner at position (x, y) in the first image, the search for a match considers all corners within a region centered on (x, y) in the second image with a threshold on maximum disparity. The s... |

1250 | Bayesian Data Analysis - Gelman, Carlin, et al. - 1995 |

1094 |
Robust regression and outlier detection
- Rousseeuw, Leroy
- 1987
(Show Context)
Citation Context ... this is usually computationally infeasible, so an important question is how many subsamples of the data set are required for statistical significance. Fischler and Bolles [4] and Rousseeuw and Leroy =-=[22]-=- proposed slightly different means of calculation, but each proposition gives broadly similar numbers. Here we follow the latter’s approach. The number m of samples is chosen sufficiently high to give... |

896 | Shape and motion from image streams under orthography: a factorization method
- Tomasi, Kanade
- 1992
(Show Context)
Citation Context ...in any form reserved. 138sESTIMATING IMAGE GEOMETRY 139 for example, epipolar geometry and projectivities. These image relations are used for several purposes: (a) matching, (b) recovery of structure =-=[1, 8, 11, 27, 40]-=- (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection [14, 35, 37]. The paper is organized as follows. In Section 2 the matrix representation of the two view relatio... |

504 |
What can be seen in three dimensions with an uncalibrated stereo rig
- Faugeras
- 1992
(Show Context)
Citation Context ...ise. 2. THE TWO-VIEW RELATIONS Within this section the possible relations on the motion of points between two views are summarized, three examples are considered in detail: (a) the fundamental matrix =-=[3, 10]-=-, (b) the planar projective transformation (projectivity), and (c) the quadratic transformation. All these two view relations are estimable from image correspondences alone. The epipolar constraint is... |

475 | A robust technique for matching two uncalibrated images through the recovery of the unknown epipolar geometry. Artif Intell J 78: 87–119 Frédéric Devernay graduated from Ecole Polytechnique (Palaiseau, France) in 1992. He obtained his Ph.D. in 1997, direc
- Zhang, Deriche, et al.
- 1995
(Show Context)
Citation Context ...using a robust technique. In this case “outliers” are putative “matches” inconsistent with the epipolar geometry. Robust estimation by random sampling (such as RANSAC ) has proven the most successful =-=[4, 29, 39]-=-. First we describe the application of RANSAC to the estimation of the fundamental matrix. Putative fundamental matrices (up to three real solutions) are computed from random sets of seven corner corr... |

466 |
Mixture Models: Inference and Applications to Clustering
- McLachlan, Basford
- 1988
(Show Context)
Citation Context ... + ˆy¯ − y i i i j=1,2 (2σ 2 ) + (1 − γ ) 1 v �� . (10) Given a suitable initial estimate there are several ways to estimate the parameters of the mixture model, most prominent being the EM algorithm =-=[2, 20]-=-, but gradient descent methods could also be used. Because of the presence of outliers in the data the standard method of least-squares estimation is often not suitable as an initial estimate, and it ... |

440 |
The Advanced Theory of Statistics
- Kendall, Stuart
- 1961
(Show Context)
Citation Context ...Given a set of correspondences the (unbiased) minimum variance solution for F is that which minimizes the sum of squares of distances orthogonal to the variety from each point (x, y, x ′ , y ′ )inR 4 =-=[12, 14, 15, 21, 23, 26, 35]-=-. This is directly equivalent to the reprojection error of the back-projected 3D projective point. In previous work such as [17] the transfer error has often been used as the error function e.g., for ... |

277 | Estimation of Relative Camera Positions for Uncalibrated Cameras
- Hartley
- 1992
(Show Context)
Citation Context ...ise. 2. THE TWO-VIEW RELATIONS Within this section the possible relations on the motion of points between two views are summarized, three examples are considered in detail: (a) the fundamental matrix =-=[3, 10]-=-, (b) the planar projective transformation (projectivity), and (c) the quadratic transformation. All these two view relations are estimable from image correspondences alone. The epipolar constraint is... |

242 |
Statistical Optimization for Geometric Computation: Theory and Practice
- Kanatani
- 1996
(Show Context)
Citation Context ...These image relations are used for several purposes: (a) matching, (b) recovery of structure [1, 8, 11, 27, 40] (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection =-=[14, 35, 37]-=-. The paper is organized as follows. In Section 2 the matrix representation of the two view relations are given, including the constraints that the matrix elements must satisfy. For example, there is ... |

236 | Euclidean reconstruction from uncalibrated views
- Hartley
(Show Context)
Citation Context ...in any form reserved. 138sESTIMATING IMAGE GEOMETRY 139 for example, epipolar geometry and projectivities. These image relations are used for several purposes: (a) matching, (b) recovery of structure =-=[1, 8, 11, 27, 40]-=- (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection [14, 35, 37]. The paper is organized as follows. In Section 2 the matrix representation of the two view relatio... |

233 | Projection pursuit - Huber - 1985 |

227 |
Estimation of planar curves, surfaces and nonplanar space curves defined by implicit equations, with applications to edge and range image segmentation
- Taubin
- 1991
(Show Context)
Citation Context ...Given a set of correspondences the (unbiased) minimum variance solution for F is that which minimizes the sum of squares of distances orthogonal to the variety from each point (x, y, x ′ , y ′ )inR 4 =-=[12, 14, 15, 21, 23, 26, 35]-=-. This is directly equivalent to the reprojection error of the back-projected 3D projective point. In previous work such as [17] the transfer error has often been used as the error function e.g., for ... |

220 | The development and comparison of robust methods for estimating the fundamental matrix
- Torr, Murray
- 1997
(Show Context)
Citation Context ...ctivity). Hartley and Sturm [12] show how e, ˆx, and ˆx ′ may be found as the solution of a degree 6 polynomial. A computationally efficient first-order approximation to these is given in Torr et al. =-=[31, 34, 35]-=-. The above derivation assumes that the errors are Gaussian; often, however, features are mismatched and the error on m is not Gaussian. Thus the error is modeled as a mixture model of the Gaussian an... |

111 | Direct least-squares fitting of algebraic surfaces
- Pratt
- 1987
(Show Context)
Citation Context ...Given a set of correspondences the (unbiased) minimum variance solution for F is that which minimizes the sum of squares of distances orthogonal to the variety from each point (x, y, x ′ , y ′ )inR 4 =-=[12, 14, 15, 21, 23, 26, 35]-=-. This is directly equivalent to the reprojection error of the back-projected 3D projective point. In previous work such as [17] the transfer error has often been used as the error function e.g., for ... |

108 | An integrated Bayesian approach to layer extraction from image sequences - Torr, Szeliski, et al. - 2001 |

103 |
Fitting Conic Sections to Very Scattered Data: An Iterative Refinement of the Bookstein Algorithm
- Sampson
- 1982
(Show Context)
Citation Context |

101 | Outlier detection and motion segmentation
- Torr, Murray
(Show Context)
Citation Context ...using a robust technique. In this case “outliers” are putative “matches” inconsistent with the epipolar geometry. Robust estimation by random sampling (such as RANSAC ) has proven the most successful =-=[4, 29, 39]-=-. First we describe the application of RANSAC to the estimation of the fundamental matrix. Putative fundamental matrices (up to three real solutions) are computed from random sets of seven corner corr... |

99 | Robust parameterization and computation of the trifocal tensor
- Torr, Zisserman
- 1997
(Show Context)
Citation Context ... (Sometimes the seven points may provide three solutions, in which case the one with lowest error is used.) This method of parameterization in terms of points was first proposed in Torr and Zisserman =-=[33]-=-.sESTIMATING IMAGE GEOMETRY 147 A number of variations on the free/fixed partition will now be discussed, as well as constraints on the direction of movement during the minimization. In all cases the ... |

92 | On determining the fundamental matrix: Analysis of different methods and experimental results - Luong, Deriche, et al. - 1993 |

69 | The geometry of projective reconstruction I: Matching constraints and the joint image
- Triggs
- 1995
(Show Context)
Citation Context ... Image Space ′ Each pair of corresponding points x, x defines a single point in a measurement space ¯ ¯ R4 , formed by considering the coordinates in each image. This space is the “joint image space” =-=[38]-=- or the “concatenated image space” [24]. It might be considered somewhat eldritch to join the coordinates of the two images into the same space, but this makes sense if we assume that the data are per... |

66 |
3D Dynamic Scene Analysis
- Zhang, Faugeras
- 1992
(Show Context)
Citation Context ...in any form reserved. 138sESTIMATING IMAGE GEOMETRY 139 for example, epipolar geometry and projectivities. These image relations are used for several purposes: (a) matching, (b) recovery of structure =-=[1, 8, 11, 27, 40]-=- (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection [14, 35, 37]. The paper is organized as follows. In Section 2 the matrix representation of the two view relatio... |

65 |
Affine Analysis of Image Sequences
- Shapiro
- 1993
(Show Context)
Citation Context ...ng points x, x defines a single point in a measurement space ¯ ¯ R4 , formed by considering the coordinates in each image. This space is the “joint image space” [38] or the “concatenated image space” =-=[24]-=-. It might be considered somewhat eldritch to join the coordinates of the two images into the same space, but this makes sense if we assume that the data are perturbed by the same noise model (discuss... |

53 | An assessment of information criteria for motion model selection
- Torr
- 1997
(Show Context)
Citation Context ...These image relations are used for several purposes: (a) matching, (b) recovery of structure [1, 8, 11, 27, 40] (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection =-=[14, 35, 37]-=-. The paper is organized as follows. In Section 2 the matrix representation of the two view relations are given, including the constraints that the matrix elements must satisfy. For example, there is ... |

52 | Robust computation and parametrization of multiple view relations - Torr, Zisserman - 1998 |

38 |
Algorithms for the solution of non-linear least squares problem
- Gill, Murray
- 1978
(Show Context)
Citation Context ...es and epipolar homography designed by Luong et al. [16]; it is both minimal and consistent. After MLESAC is applied, nonlinear minimization is conducted using the method described in Gill and Murray =-=[6]-=-, which is a modification of the Gauss–Newton method. All the points are included in the minimization, but the effect of outliers is removed as the robust function places a ceiling on the value of the... |

38 | Determining the Fundamental matrix with planes: unstability and new algorithms
- Luong, Faugeras
- 1993
(Show Context)
Citation Context ...variety from each point (x, y, x ′ , y ′ )inR 4 [12, 14, 15, 21, 23, 26, 35]. This is directly equivalent to the reprojection error of the back-projected 3D projective point. In previous work such as =-=[17]-=- the transfer error has often been used as the error function e.g., for fitting H. This is d 2 (x, H −1 x ′ ) + d 2 (x ′ , Hx), (5) where d( ) is the Euclidean image distance between the points. The t... |

33 | Robust detection of Degenerate Configurations for the fundamental matrix
- Torr, Zisserman
- 1995
(Show Context)
Citation Context ...These image relations are used for several purposes: (a) matching, (b) recovery of structure [1, 8, 11, 27, 40] (if this is possible), (c) motion segmentation [30, 36], and (d) motion model selection =-=[14, 35, 37]-=-. The paper is organized as follows. In Section 2 the matrix representation of the two view relations are given, including the constraints that the matrix elements must satisfy. For example, there is ... |

31 |
3D model aquisition from extended image sequence
- Beardsley, Zisserman
- 1996
(Show Context)
Citation Context |

31 | Bias in robust estimation caused by discontinuities and multiple structures
- Stewart
- 1997
(Show Context)
Citation Context ...ional burden. Once this is understood there is no reason to use RANSAC in preference to this method. Similar schemes for robust estimation using random sampling and M-estimators were also proposed in =-=[25, 28]-=-. The definition of the maximum likelihood error allows us to suggest a further improvement over MSAC. As the aim is to minimize the negative log likelihood of the mixture −L, it makes sense to use th... |

30 |
Robust Computation and Parameterization of Multiple View Relations
- Torr, Zisserman
- 1998
(Show Context)
Citation Context ...ctivity). Hartley and Sturm [12] show how e, ˆx, and ˆx ′ may be found as the solution of a degree 6 polynomial. A computationally efficient first-order approximation to these is given in Torr et al. =-=[31, 34, 35]-=-. The above derivation assumes that the errors are Gaussian; often, however, features are mismatched and the error on m is not Gaussian. Thus the error is modeled as a mixture model of the Gaussian an... |

29 |
Stochastic motion clustering
- Torr, Murray
- 1994
(Show Context)
Citation Context ...e, epipolar geometry and projectivities. These image relations are used for several purposes: (a) matching, (b) recovery of structure [1, 8, 11, 27, 40] (if this is possible), (c) motion segmentation =-=[30, 36]-=-, and (d) motion model selection [14, 35, 37]. The paper is organized as follows. In Section 2 the matrix representation of the two view relations are given, including the constraints that the matrix ... |

26 | Motion clustering using the trilinear constraint over three views
- Torr, Zisserman, et al.
- 1995
(Show Context)
Citation Context ...e, epipolar geometry and projectivities. These image relations are used for several purposes: (a) matching, (b) recovery of structure [1, 8, 11, 27, 40] (if this is possible), (c) motion segmentation =-=[30, 36]-=-, and (d) motion model selection [14, 35, 37]. The paper is organized as follows. In Section 2 the matrix representation of the two view relations are given, including the constraints that the matrix ... |

20 | and O.D.Faugeras. A stability analysis for the fundamental matrix
- Luong
- 1994
(Show Context)
Citation Context ...nstraint is imposed. The projectivity has 9 elements and 8 degrees of freedom, as these elements are only defined up to a scale. The quadratic transformation has 18 elements and 14 degrees of freedom =-=[18]-=-. Here if the constraints between the parameters are not enforced the estimation process becomes very unstable, and good results cannot be obtained [18], whereas our method has been able to accurately... |

20 |
Robust vision
- Torr, Beardsley, et al.
- 1994
(Show Context)
Citation Context ...ional burden. Once this is understood there is no reason to use RANSAC in preference to this method. Similar schemes for robust estimation using random sampling and M-estimators were also proposed in =-=[25, 28]-=-. The definition of the maximum likelihood error allows us to suggest a further improvement over MSAC. As the aim is to minimize the negative log likelihood of the mixture −L, it makes sense to use th... |

11 |
Properties of Essential Matrices
- Maybank
- 1990
(Show Context)
Citation Context ...nsformation is a generalization of the homography. It is caused by a combination of a camera motion and scene structure, as all the scene points and the camera optic centers lie on a critical surface =-=[19]-=-, which is a ruled quadric surface. Although the existence of the critical surface is well known, little research has been put into effectively estimating quadratic transformations. 2.1. Degrees of Fr... |

8 |
The Droid 3D Vision System
- Harris
- 1988
(Show Context)
Citation Context |

8 |
On determining the fundamental matrix
- Luong, Deriche, et al.
- 1993
(Show Context)
Citation Context ...e not independent, being related by a cubic polynomial in the matrix elements, such that det[F] = 0. If this constraint is not imposed then the epipolar lines do not all intersect in a single epipole =-=[16]-=-. Hence it is essential that this constraint is imposed. The projectivity has 9 elements and 8 degrees of freedom, as these elements are only defined up to a scale. The quadratic transformation has 18... |

1 |
An integrated approach to layer extraction
- Torr, Szeliski, et al.
- 1999
(Show Context)
Citation Context ...hod may be generalized to the case when the data has arisen from a more general mixture model involving several classes, such as in clustering problems. Preliminary work to illustrate this appears in =-=[32]-=-. There are two trivial extensions to MLESAC that allow the introduction of prior information, although prior information is not used in this paper. The first is to allow a prior Pr(M) on the paramete... |