#### DMCA

## Deformation-Aware Log-Linear Models

### Cached

### Download Links

Citations: | 2 - 2 self |

### Citations

3485 | Conditional random fields: Probabilistic models for segmenting and labeling sequence datasets
- Lafferty, McCallum, et al.
- 2001
(Show Context)
Citation Context ...M) [4], a zero-order, non-linear deformation model, which we shortly describe in the following section. The developed model can also be considered a grid-shaped hidden-conditional random field (HCRF) =-=[7, 8]-=- where the latent variables account for the deformations. In section 3, we present our model which incorporates the IDM into log-linear models. In section 4, we present an experimental evaluation on t... |

970 | A fast learning algorithm for deep belief nets.
- Hinton, Osindero, et al.
- 2006
(Show Context)
Citation Context ...7 040 000 3.1 729/6 000 nearest neighbor + IDM [4] 1 866 496 2.4 47 040 000 0.6 36 455/300 000 SVM 658 177 4.4 15 411 905 1.5 256/1 963 SVM + IDM [16]/[this work] 530 705 2.8 - 0.7 10 300/100 000 DBN =-=[17]-=- 640 610 - 1 665 010 1.3 210/ 220 conv. network [3] - - 180 580 0.4 -/25 MNIST dataset and compare the results for both datasets with several state-ofthe-art results from the literature. Additionally ... |

938 | A direct adaptive method for faster backpropagation learning: the RPROP algorithm
- Riedmiller, Braun
- 1993
(Show Context)
Citation Context ...on in numerator and denominator, the same procedure can be used, but here for each training observation, an alignment for each class has to be determined. We train our model using the RProp-algorithm =-=[13]-=-, which has the advantage that it is robust w.r.t. varying scales of the derivatives because it only takes into account the sign of the partial derivatives to determine the parameter updates. 3.4 Pool... |

201 | Best practices for convolutional neural networks applied to visual document analysis.
- Simard, Steinkraus, et al.
- 2003
(Show Context)
Citation Context ...meworks [4, 5]. Another approach is not to incorporate the deformation-invariance into the model but use a huge amount of synthetically deformed data during training of a convolutional neural network =-=[3]-=-. The first approach has the disadvantage that during testing a large amount of, potentially computationally expensive, image comparisons have to be performed whereas in the second approach the traini... |

186 | Training invariant support vector machines.
- Decoste, Schoelkopf
- 2002
(Show Context)
Citation Context ...s into their classification framework, e.g. incorporate the tangent distance into support vector machines [2], use kerneljittering to obtain translated support vectors in a two-step training approach =-=[1]-=- and use transformation-invariant distance measures in nearest neighbor frameworks [4, 5]. Another approach is not to incorporate the deformation-invariance into the model but use a huge amount of syn... |

122 |
Hidden conditional random fields.
- Quattoni, Wang, et al.
- 2007
(Show Context)
Citation Context ...M) [4], a zero-order, non-linear deformation model, which we shortly describe in the following section. The developed model can also be considered a grid-shaped hidden-conditional random field (HCRF) =-=[7, 8]-=- where the latent variables account for the deformations. In section 3, we present our model which incorporates the IDM into log-linear models. In section 4, we present an experimental evaluation on t... |

54 | Deformation models for image recognition.
- Keysers, Deselaers, et al.
- 2007
(Show Context)
Citation Context ...ort vector machines [2], use kerneljittering to obtain translated support vectors in a two-step training approach [1] and use transformation-invariant distance measures in nearest neighbor frameworks =-=[4, 5]-=-. Another approach is not to incorporate the deformation-invariance into the model but use a huge amount of synthetically deformed data during training of a convolutional neural network [3]. The first... |

54 |
Research on machine recognition of handprinted characters,"
- Mori, Yamamoto, et al.
- 1984
(Show Context)
Citation Context ...rnel. 2 Image Distortion Model The IDM has been proposed by several works independently under different names. For example, it has been described as “local pertubations” [9] and as “shift similarity” =-=[10]-=-. Here, we follow the formulation of [4]. The IDM is a zero order image deformation method that accounts for image transformations by pixel-wise aligning a test image to a prototype image without cons... |

50 | Unsupervised learning of image transformations
- Memisevic, Hinton
- 2007
(Show Context)
Citation Context ...ameters of the allowed deformations but the deformation model was hand-coded by the system developers.In contrast to these approaches to transformation-invariant classification, Memisevic and Hinton =-=[6]-=- proposed an approach to learn image transformations from corresponding image pairs using conditional restricted Boltzmann machines. In our approach, we aim at training a small (in the number of param... |

41 | Tangent distance kernels for support vector machines,
- Haasdonk, Keysers
- 2002
(Show Context)
Citation Context ...proaches can be split into two groups: Approaches that directly incorporate certain invariances into their classification framework, e.g. incorporate the tangent distance into support vector machines =-=[2]-=-, use kerneljittering to obtain translated support vectors in a two-step training approach [1] and use transformation-invariant distance measures in nearest neighbor frameworks [4, 5]. Another approac... |

27 |
Adaptation in statistical pattern recognition using tangent vectors,
- Keysers, Mcherey, et al.
- 2004
(Show Context)
Citation Context ... p(c, (xy) IJ 11 |X) = ∑ c ′ p(c′ ) p(X, (xy) IJ 11 |c′ ) p(X, (xy) IJ 11 |c) = p((xy) IJ 11 |c) p(X|c, (xy) IJ where p((xy) IJ 11 |c) can be considered as a deformation prior and p(X|c, (xy) IJ 11 ) =-=(5)-=- 11 ) is an emission probability for a given class and alignment. Then, p((xy) IJ 11 |c) can be rewritten as p((xy) IJ 11 |c) = ∏ ij p((xy)ij|ij, c) and p(X|c, (xy) IJ 1 11 ) = √ 2πσ2 exp ( − 1 ∑ (Xij... |

19 | Maximum Entropy and Gaussian Models for Image Object
- Keysers, Och, et al.
- 2002
(Show Context)
Citation Context ...the λc(xy)ij are of the same dimensionality. 3.1 Relationship to Gaussian Models An interesting aspect of this model is that it can be rewritten as a discriminative Gaussian classifier analogously to =-=[11]-=-. We rewrite p(c) p(X, (xy) IJ 11 |c) and decompose p(c, (xy) IJ 11 |X) = ∑ c ′ p(c′ ) p(X, (xy) IJ 11 |c′ ) p(X, (xy) IJ 11 |c) = p((xy) IJ 11 |c) p(X|c, (xy) IJ where p((xy) IJ 11 |c) can be conside... |

18 |
A survey of elastic matching techniques for handwritten character recognition”,
- Uchida, Sakoe
- 2005
(Show Context)
Citation Context ...an SVM with an IDM-distance kernel. 2 Image Distortion Model The IDM has been proposed by several works independently under different names. For example, it has been described as “local pertubations” =-=[9]-=- and as “shift similarity” [10]. Here, we follow the formulation of [4]. The IDM is a zero order image deformation method that accounts for image transformations by pixel-wise aligning a test image to... |

11 |
Transformation Knowledge in Pattern Analysis with Kernel Methods-Distance and Integration Kernels
- Haasdonk
- 2005
(Show Context)
Citation Context ... 6.5 7 840 5.8 50 nearest neighbor [4] 1 866 496 5.6 47 040 000 3.1 729/6 000 nearest neighbor + IDM [4] 1 866 496 2.4 47 040 000 0.6 36 455/300 000 SVM 658 177 4.4 15 411 905 1.5 256/1 963 SVM + IDM =-=[16]-=-/[this work] 530 705 2.8 - 0.7 10 300/100 000 DBN [17] 640 610 - 1 665 010 1.3 210/ 220 conv. network [3] - - 180 580 0.4 -/25 MNIST dataset and compare the results for both datasets with several stat... |

1 |
H.: GIS-like estimation of log-linear models with hidden variables
- Heigold, Deselaers, et al.
- 2008
(Show Context)
Citation Context ...ranteed, however, as we found in our experiments, the training converges well. An extension of the GIS algorithm to allow for training of log-linear models with hidden variables has been presented in =-=[12]-=-. However the authors observed that although the algorithm is guaranteed to converge, convergence can be slow. Similarly to their approach, we also use an alternating optimization method: Step 1: Trai... |