## Linear Fitting with Missing Data for Structure-from-Motion (1997)

Venue: | Computer Vision and Image Understanding |

Citations: | 47 - 6 self |

### BibTeX

@ARTICLE{Jacobs97linearfitting,

author = {David W. Jacobs},

title = {Linear Fitting with Missing Data for Structure-from-Motion},

journal = {Computer Vision and Image Understanding},

year = {1997},

volume = {82},

pages = {2001}

}

### Years of Citing Articles

### OpenURL

### Abstract

this paper. This method is described in detail in [15]. We can briefly describe the method as formulating the least squares problem as a bilinear optimization, and then iteratively holding one set of variables constant while the others are optimized, so that each optimization is linear. We use their method in our experiments, because it has good convergence properties and is easy to implement. For the problem they consider, Shum et al. state that a random starting point is sufficient to produce a good final solution. However, their experiments on this point cannot be used to draw conclusions for the problem of determining 3-D structure from a sequence of 2-D images. 3 A Novel Algorithm

### Citations

1316 |
Statistical Analysis with Missing Data
- LITTLE, RUBIN
- 1987
(Show Context)
Citation Context ...uch as EM are used to find maximum likelihood estimates for unknown values, given some model of their properties. A taxonomy of these methods and a detailed discussion can be found in Little and Rubin=-=[8]-=-. Several methods have been suggested in computer vision that are suitable for computing structure-from-motion with missing data. As in many non-linear optimization problems, solutions can generally b... |

930 | T.: Shape and motion from image streams under orthography: A factorization method
- Tomasi, Kanade
- 1992
(Show Context)
Citation Context ...ed by any 3-D structure from different viewpoints is represented as a low-dimensional surface in the space of all possible images. This structure is then used for object recognition. Tomasi and Kanade=-=[18]-=- use a related insight to efficiently perform structure-from-motion in long motion sequences. Since the noise-free images should all lie in a low-dimensional linear surface, such a surface can be fit ... |

423 |
R.: Recognition by linear combinations of models
- Ullman, Basri
- 1991
(Show Context)
Citation Context ...g models, which reduce the problems to ones of finding a low-dimensional linear surface that represents some important scene structure. For example, in the linear combinations work of Basri and Ullman=-=[19, 2]-=-, the set of images produced by any 3-D structure from different viewpoints is represented as a low-dimensional surface in the space of all possible images. This structure is then used for object reco... |

337 | What is the set of images of an object under all possible lighting conditions
- Belhumeur, Kriegman
- 1996
(Show Context)
Citation Context ...an structure from a single viewpoint but with a point light source of varying position and intensity can also be described as a 3-D linear surface in a space of possible images. Belhumeur and Kriegman=-=[3]-=- use this result to describe the set of intensity images produced by a convex object, allowing for multiple light sources and self-shadowing. In all of this work, the problem is greatly simplified by ... |

257 | A Paraperspective Factorization Method for Shape and Motion Recovery, tech
- Poelman, Kanade
- 1993
(Show Context)
Citation Context ...ence of matched image points is given, and that these were projected from 3-D to 2-D using a camera model that is affine, or a subset of affine (eg. orthographic, weak perspective, or paraperspective =-=[11, 17, 1]-=-). The translation component of motion is estimated separately, by estimating the translation of the center of mass of the points. This method is no longer suitable when there is missing data. To simp... |

241 | Euclidean reconstruction from uncalibrated views
- Hartley
- 1993
(Show Context)
Citation Context ... it is not guaranteed to find the globally optimal one. We describe this method in Section 2, and later describe experiments in which we apply their method to our problems. Other authors (eg., Hartley=-=[6]-=- McLauchlan et al.[9], and Soatto and Perona[16]) have also applied iterative error minimization methods to the problem of determining structure-from-motion. We show, as others have also noted (eg., H... |

102 |
Principal Component Analysis with Missing Data and Its Application to Polyhedral Object Modeling
- Shum, Reddy
- 1995
(Show Context)
Citation Context ...ank matrix to a data matrix that has some missing elements, and provided a method of doing this. We describe this method in Section 2, and compare it to our own in Section 4.2. Shum, Ikeuchi and Reddy=-=[15]-=- apply an iterative method due to Wiberg[20] to a related structure from motion problem. This method minimizes the sum of square differences between the fitted, low rank matrix and the elements that a... |

62 |
fitting with missing data: Applications to structure-from-motion and to characterizing intensity images
- Linear
- 1997
(Show Context)
Citation Context ...ons when there is less data present. We also show experimentally that our method provides good results compared to previously proposed methods. A more preliminary version of this work has appeared as =-=[7]-=- 1 Introduction Recently, several important vision problems have been addressed by using affine imaging models, which reduce the problems to ones of finding a low-dimensional linear surface that repre... |

56 |
Checking Robust Nonsingularity Is NP-Hard
- Poljak, Rohn
- 1993
(Show Context)
Citation Context ...s do not exist for the general problem of fitting a low rank matrix to a matrix with missing elements, even in the absence of sensing error, and closely related problems are known to be NP-hard (eg., =-=[12]-=-). Related problems have also been studied in statistics. While many of the methods developed are specialized for regression analysis, two methods are related to techniques that have been used in comp... |

45 | The alignment of objects with smooth surfaces
- Basri, Ullman
- 1993
(Show Context)
Citation Context ...g models, which reduce the problems to ones of finding a low-dimensional linear surface that represents some important scene structure. For example, in the linear combinations work of Basri and Ullman=-=[19, 2]-=-, the set of images produced by any 3-D structure from different viewpoints is represented as a low-dimensional surface in the space of all possible images. This structure is then used for object reco... |

43 |
Computation of principal components when data are missing
- Wiberg
- 1976
(Show Context)
Citation Context ...ssing elements, and provided a method of doing this. We describe this method in Section 2, and compare it to our own in Section 4.2. Shum, Ikeuchi and Reddy[15] apply an iterative method due to Wiberg=-=[20]-=- to a related structure from motion problem. This method minimizes the sum of square differences between the fitted, low rank matrix and the elements that are not missing in the data matrix. This appr... |

41 | Recursive affine structure and motion from image sequences
- McLauchlan, Reid, et al.
- 1994
(Show Context)
Citation Context ... to find the globally optimal one. We describe this method in Section 2, and later describe experiments in which we apply their method to our problems. Other authors (eg., Hartley[6] McLauchlan et al.=-=[9]-=-, and Soatto and Perona[16]) have also applied iterative error minimization methods to the problem of determining structure-from-motion. We show, as others have also noted (eg., Hartley[6]) that these... |

27 |
Face recognition: generalization to novel images
- Moses
- 1993
(Show Context)
Citation Context ...tion in long motion sequences. Since the noise-free images should all lie in a low-dimensional linear surface, such a surface can be fit to the data and then used to derive structure and motion. Moses=-=[10]-=- and Shashua[14] show that the set of intensity images produced by a 3-D Lambertian structure from a single viewpoint but with a point light source of varying position and intensity can also be descri... |

24 | Geometry and Photometry
- Shashua
- 1992
(Show Context)
Citation Context ...ion sequences. Since the noise-free images should all lie in a low-dimensional linear surface, such a surface can be fit to the data and then used to derive structure and motion. Moses[10] and Shashua=-=[14]-=- show that the set of intensity images produced by a 3-D Lambertian structure from a single viewpoint but with a point light source of varying position and intensity can also be described as a 3-D lin... |

20 |
The smallest perturbation of a submatrix which lowers the rank and constrained total least squares problems
- Demmel
- 1987
(Show Context)
Citation Context ..., inspired in part by problems that arise in control. Solutions have been found for specialized versions of the problem, for example, when the missing elements form a rectangular block of the matrix (=-=[4]-=-). However, solutions do not exist for the general problem of fitting a low rank matrix to a matrix with missing elements, even in the absence of sensing error, and closely related problems are known ... |

4 |
Dynamic rigid motion estimation from weak perspective
- Soatto, Perona
- 1995
(Show Context)
Citation Context ...mal one. We describe this method in Section 2, and later describe experiments in which we apply their method to our problems. Other authors (eg., Hartley[6] McLauchlan et al.[9], and Soatto and Perona=-=[16]-=-) have also applied iterative error minimization methods to the problem of determining structure-from-motion. We show, as others have also noted (eg., Hartley[6]) that these methods can often converge... |

3 |
Description and Reconstruction from Image
- Sawhney, Oliensis, et al.
- 1990
(Show Context)
Citation Context ...ith a standard deviation of :01. Finally, intensities less than zero were considered to be missing values. Finally, we have run our algorithm on data obtained from a real motion sequence described in =-=[13]-=-. This sequence contained forty points tracked across eight images. One of the frames is shown in Figure 1. In this sequence we generated artificial occlusions exactly as described for the random moti... |

2 |
Object Recognition by Combining Paraperspective Images
- Sugihara
- 1996
(Show Context)
Citation Context ...ence of matched image points is given, and that these were projected from 3-D to 2-D using a camera model that is affine, or a subset of affine (eg. orthographic, weak perspective, or paraperspective =-=[11, 17, 1]-=-). The translation component of motion is estimated separately, by estimating the translation of the center of mass of the points. This method is no longer suitable when there is missing data. To simp... |

1 |
Paraperspective j Affine," Int
- Basri
- 1996
(Show Context)
Citation Context ...ence of matched image points is given, and that these were projected from 3-D to 2-D using a camera model that is affine, or a subset of affine (eg. orthographic, weak perspective, or paraperspective =-=[11, 17, 1]-=-). The translation component of motion is estimated separately, by estimating the translation of the center of mass of the points. This method is no longer suitable when there is missing data. To simp... |