Metric Learning from Relative Comparisons by Minimizing Squared Residual

Liu, Eric Yi; Guo, Zhishan; Xiang, Zhang; Jojic, Vladimir; Wang, Wei

doi:10.1109/icdm.2012.38

Cited by 43 publications

(38 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…By virtue of this training principle, [35] has significantly boosted the performance of kinship verification. Liu et al [33] developed a metric learning algorithm by minimizing squared residuals from relative comparisons, where pair-wise constraints were not natural to obtain. While, these two approaches could not be directly used to exploit the relative similarity information in metric learning for person re-identification, due to the unavailability of relative similarity information in practical person re-identification.…”

Section: Related Workmentioning

confidence: 99%

Relevance Metric Learning for Person Re-Identification by Exploiting Listwise Similarities

Chen

Zhang

Wang

2015

IEEE Trans. on Image Process.

View full text Add to dashboard Cite

Person re-identification aims to match people across non-overlapping camera views, which is an important but challenging task in video surveillance. In order to obtain a robust metric for matching, metric learning has been introduced recently. Most existing works focus on seeking a Mahalanobis distance by employing sparse pairwise constraints, which utilize image pairs with the same person identity as positive samples, and select a small portion of those with different identities as negative samples. However, this training strategy has abandoned a large amount of discriminative information, and ignored the relative similarities. In this paper, we propose a novel relevance metric learning method with listwise constraints (RMLLCs) by adopting listwise similarities, which consist of the similarity list of each image with respect to all remaining images. By virtue of listwise similarities, RMLLC could capture all pairwise similarities, and consequently learn a more discriminative metric by enforcing the metric to conserve predefined similarity lists in a low-dimensional projection subspace. Despite the performance enhancement, RMLLC using predefined similarity lists fails to capture the relative relevance information, which is often unavailable in practice. To address this problem, we further introduce a rectification term to automatically exploit the relative similarities, and develop an efficient alternating iterative algorithm to jointly learn the optimal metric and the rectification term. Extensive experiments on four publicly available benchmarking data sets are carried out and demonstrate that the proposed method is significantly superior to the state-of-the-art approaches. The results also show that the introduction of the rectification term could further boost the performance of RMLLC.

show abstract

Section: Related Workmentioning

confidence: 99%

Relevance Metric Learning for Person Re-Identification by Exploiting Listwise Similarities

Chen

Zhang

Wang

2015

IEEE Trans. on Image Process.

View full text Add to dashboard Cite

show abstract

“…Computing L 2 between data points is a strong baseline, which sometimes outperforms true distance learning methods, especially for high-dimensional data. Four of the distance-learning methods [36,10,26,21] learn a global linear transformation of the data, which is commonly referred to as the Mahalanobis metric learning. The value of the L 2 distance between transformed vectors is used as a proxy distance function.…”

Section: Proxying Distance Via Metrization and Symmetrizationmentioning

confidence: 99%

Accurate and Fast Retrieval for Complex Non-metric Data via Neighborhood Graphs

Boytsov

Nyberg

2019

Similarity Search and Applications

View full text Add to dashboard Cite

We demonstrate that a graph-based search algorithm-relying on the construction of an approximate neighborhood graph-can directly work with challenging non-metric and/or non-symmetric distances without resorting to metricspace mapping and/or distance symmetrization, which, in turn, lead to substantial performance degradation. Although the straightforward metrization and symmetrization is usually ineffective, we find that constructing an index using a modified, e.g., symmetrized, distance can improve performance. This observation paves a way to a new line of research of designing index-specific graph-construction distance functions. This is an archival version, the publisher's version is available at Springer.com.

show abstract

“…Depending on this performance, triplets are dynamically selected so that points near boundaries are used as constraint points. This change of triplets over the iterations gives more interesting results than the triplet selection method proposed in [23]. However, they only focus on points near boundaries; We propose to focus on points near boundaries, while also keeping also observations that are far from the boundaries.…”

Section: Metric Learning and Related Workmentioning

confidence: 99%

“…4) Relative labels [23]: are considering four different objects x i , x j , x k and x l for writing the constraint. In this setting, x i and x j are chosen so that their distance is lower than the distance between x k and x l : d(x i , x j ) ≤ d(x k , x l ).…”

Section: Dynamic Local Weights For Metric Learning a Problem Anmentioning

confidence: 99%