DMCA
H.: Large scale metric learning from equivalence constraints (2012)
| Venue: | In: Proc. IEEE Intern. Conf. on Computer Vision and Pattern Recognition |
| Citations: | 77 - 5 self |
Citations
| 8954 | Distinctive image features from scale-invariant keypoint - Lowe - 2004 |
| 6495 | LIBSVM: A Library for Support Vector Machines
- Chang, Lin
(Show Context)
Citation Context ... based approach proposed by [13]. the face descriptors as this delivers better results. In Figure 2 (b) we report a Receiver Operator Characteristic (ROC) curve for LDML [8], ITML [3], LMNN [19], SVM =-=[2]-=-, our method (KISSME), the Mahalanobis distance of the similar pairs and the Euclidean distance as baseline. Please note that for LMNN we have to provide more supervision in form of the actual class l... |
| 1299 | Multiresolution gray-scale and rotation invariant texture classification with local binary patterns - Ojala, Pietikäinen, et al. - 2002 |
| 695 | Distance metric learning for large margin nearest neighbor classification
- Weinberger, Saul
(Show Context)
Citation Context ...ance is measured. Machine learning algorithms that learn a Mahalanobis metric have recently attracted a lot of interest in computer vision. These include Large Margin Nearest Neighbor Learning (LMNN) =-=[19, 20]-=-, Information Theoretic Metric Learning (ITML) [3] and Logistic Discriminant Metric Learning (LDML) [8], which can be considered as stateof-the-art. LMNN [19, 20] aims at improving k-nn clas-sificati... |
| 494 |
The Relaxation Method of Finding the Common Point of Convex Sets and its Application to the Solution
- Bregman
- 1967
(Show Context)
Citation Context ...onstraints enforce that similar pairs are below a certain distance d2 M (xi, xj) ≤ u while dissimilar pairs exceed a certain distance d2 M (xi, xj) ≥ l. The optimization builds on Bregman projections =-=[1]-=-, which project the current solution onto a single constraint via the update rule: Mt+1 = Mt + βMtCijMt . (5) The parameter β involves the pair label and the step size. It is positive for similar pair... |
| 449 | Labeled faces in the wild: A database for studying face recognition in unconstrained environments.
- Huang, Ramesh, et al.
- 2007
(Show Context)
Citation Context ...tion Agency (FFG) project SHARE in the IV2Splus program. TP FP FN TN (a) Distribution of distances before . . . TP FP FN TN (b) . . . and after applying our method. Figure 1: Face verification on LFW =-=[12]-=-: The challenging task shows the benefit of metric learning. Our method significantly increases the TPR at EER from 67.4% (a) to 80.5% (b). Training takes only 0.05 seconds and is thus orders of magni... |
| 359 | Informationtheoretic metric learning
- Davis, Kulis, et al.
- 2007
(Show Context)
Citation Context ...n a Mahalanobis metric have recently attracted a lot of interest in computer vision. These include Large Margin Nearest Neighbor Learning (LMNN) [19, 20], Information Theoretic Metric Learning (ITML) =-=[3]-=- and Logistic Discriminant Metric Learning (LDML) [8], which can be considered as stateof-the-art. LMNN [19, 20] aims at improving k-nn clas-sification. It establishes for each instance a local perim... |
| 325 | Attribute and simile classifiers for face verification.
- Kumar, Berg, et al.
- 2009
(Show Context)
Citation Context ...tperform state-of-the-art metric learning approaches, while being orders of magnitudes faster in training. In particular, we provide results on two recent face recognition benchmarks (LFW [12], PubFig=-=[13]-=-). Due to the non-rigid nature and changes in pose, lighting and expression faces are a challenge for learning algorithms. Further, we study the task of person reidentification across spatially disjoi... |
| 159 | Is that you? Metric learning approaches for face identification. In
- Guillaumin, Verbeek, et al.
- 2009
(Show Context)
Citation Context ...of interest in computer vision. These include Large Margin Nearest Neighbor Learning (LMNN) [19, 20], Information Theoretic Metric Learning (ITML) [3] and Logistic Discriminant Metric Learning (LDML) =-=[8]-=-, which can be considered as stateof-the-art. LMNN [19, 20] aims at improving k-nn clas-sification. It establishes for each instance a local perimeter. The perimeter surrounds the k-nns with similar ... |
| 152 | Person re-identification by symmetry-driven accumulation of local features. In
- Farenzena, Bazzani, et al.
- 2010
(Show Context)
Citation Context ... about 90 degrees, making person re-identification very challenging. Some examples are given in Figure 3 (a). To compare our method to other approaches, we followed the evaluation protocol defined in =-=[5, 7]-=-. The authors split the set of 632 image pairs randomly into two sets of 316 image pairs each, one for training and one for testing, and compute the average over several runs. There is no predefined s... |
| 141 |
Viewpoint invariant pedestrian recognition with an ensemble of localized features.
- Gray, Tao
- 2008
(Show Context)
Citation Context ... about 90 degrees, making person re-identification very challenging. Some examples are given in Figure 3 (a). To compare our method to other approaches, we followed the evaluation protocol defined in =-=[5, 7]-=-. The authors split the set of 632 image pairs randomly into two sets of 316 image pairs each, one for training and one for testing, and compute the average over several runs. There is no predefined s... |
| 103 | Person reidentification by support vector ranking.
- Prosser, Zheng, et al.
- 2010
(Show Context)
Citation Context ...formance of our approach in the range of the first 50 ranks to state-of-the-art methods [4, 5, 10, 23]. As can be seen, we obtain competitive results across all ranks. We outperform the other methods =-=[5, 7, 17]-=- even though in contrastRANK 1 10 25 50 KISSME 19.6% 62.2% 80.7% 91.8% LMNN 19.0% 58.1% 76.9% 89.6% ITML 15.2% 53.3% 74.7% 88.8% LDML 10.4% 31.3% 44.6% 60.4% My=1 16.8% 50.9% 68.7% 82.0% L2 10.6% 31.... |
| 93 | Shape and appearance context modeling. In
- Wang, Doretto, et al.
- 2007
(Show Context)
Citation Context ...proaches we project the concatenated descriptors into a 34 dimensional subspace by PCA. To indicate the performance of the various algorithms we report Cumulative Matching Characteristic (CMC) curves =-=[18]-=-. These represent the expectation of the true match being found within the first n ranks. To obtain a reasonable statistical significance, we average over 100 runs. In Figure 3 (b) we report the CMC c... |
| 85 | Fast solvers and efficient implementations for distance metric learning.
- Weinberger, Saul
- 2008
(Show Context)
Citation Context ...ance is measured. Machine learning algorithms that learn a Mahalanobis metric have recently attracted a lot of interest in computer vision. These include Large Margin Nearest Neighbor Learning (LMNN) =-=[19, 20]-=-, Information Theoretic Metric Learning (ITML) [3] and Logistic Discriminant Metric Learning (LDML) [8], which can be considered as stateof-the-art. LMNN [19, 20] aims at improving k-nn clas-sificati... |
| 82 | Jurie F.: Learning visual similarity measures for comparing never seen objects
- Nowak
(Show Context)
Citation Context ...llenge for learning algorithms. Further, we study the task of person reidentification across spatially disjoint cameras (VIPeR [6]) and the comparison of before never seen object instances on ToyCars =-=[15]-=-. On VIPeR and the ToyCars dataset we improve even over the domain specific state-of-the-art. Further, for LFW we obtain the best reported results for standard SIFT features. The rest of this paper is... |
| 81 | W.Y.: Learning distance metrics with contextual constraints for image retrieval. CVPR
- Hoi, Liu, et al.
- 2006
(Show Context)
Citation Context ...Learning distance or similarity metrics is an emerging field in machine learning, with various applications in computer vision. It can significantly improve results for tracking [22], image retrieval =-=[11]-=-, face identification [9], clustering [21], or person re-identification [4]. The goal of metric learning algorithms is to take advantage of prior information in form of labels over simpler though more... |
| 63 | Person re-identification by probabilistic relative distance comparison.
- Zheng, Gong, et al.
- 2011
(Show Context)
Citation Context ...we report the CMC curves for the various metric learning algorithms. Moreover, in Table 2 (b) we compare the performance of our approach in the range of the first 50 ranks to state-of-the-art methods =-=[4, 5, 10, 23]-=-. As can be seen, we obtain competitive results across all ranks. We outperform the other methods [5, 7, 17] even though in contrastRANK 1 10 25 50 KISSME 19.6% 62.2% 80.7% 91.8% LMNN 19.0% 58.1% 76.... |
| 59 | Pedestrian recognition with a learned metric. In ACCV.
- Dikmen, Akbas, et al.
- 2011
(Show Context)
Citation Context ...we report the CMC curves for the various metric learning algorithms. Moreover, in Table 2 (b) we compare the performance of our approach in the range of the first 50 ranks to state-of-the-art methods =-=[4, 5, 10, 23]-=-. As can be seen, we obtain competitive results across all ranks. We outperform the other methods [5, 7, 17] even though in contrastRANK 1 10 25 50 KISSME 19.6% 62.2% 80.7% 91.8% LMNN 19.0% 58.1% 76.... |
| 44 | h.: Person re-identification by descriptive and discriminative classification.
- hirzer, Beleznai, et al.
- 2011
(Show Context)
Citation Context ...we report the CMC curves for the various metric learning algorithms. Moreover, in Table 2 (b) we compare the performance of our approach in the range of the first 50 ranks to state-of-the-art methods =-=[4, 5, 10, 23]-=-. As can be seen, we obtain competitive results across all ranks. We outperform the other methods [5, 7, 17] even though in contrastRANK 1 10 25 50 KISSME 19.6% 62.2% 80.7% 91.8% LMNN 19.0% 58.1% 76.... |
| 31 | Multiple instance metric learning from automatically labeled bags of faces. - Guillaumin, Verbeek, et al. - 2010 |
| 22 |
Adaptive distance metric learning for clustering
- Ye, Zhao, et al.
- 2007
(Show Context)
Citation Context ... an emerging field in machine learning, with various applications in computer vision. It can significantly improve results for tracking [22], image retrieval [11], face identification [9], clustering =-=[21]-=-, or person re-identification [4]. The goal of metric learning algorithms is to take advantage of prior information in form of labels over simpler though more general similarity measures, illustrated ... |
| 21 | Toward robust distance metric analysis for similarity estimation.
- Yu, Amores, et al.
- 2006
(Show Context)
Citation Context ...-art. 1. Introduction Learning distance or similarity metrics is an emerging field in machine learning, with various applications in computer vision. It can significantly improve results for tracking =-=[22]-=-, image retrieval [11], face identification [9], clustering [21], or person re-identification [4]. The goal of metric learning algorithms is to take advantage of prior information in form of labels ov... |
| 11 |
Evaluating appearance models for recongnition, reacquisition and tracking
- Gray, Brennan, et al.
(Show Context)
Citation Context ...rigid nature and changes in pose, lighting and expression faces are a challenge for learning algorithms. Further, we study the task of person reidentification across spatially disjoint cameras (VIPeR =-=[6]-=-) and the comparison of before never seen object instances on ToyCars [15]. On VIPeR and the ToyCars dataset we improve even over the domain specific state-of-the-art. Further, for LFW we obtain the b... |
