Correcting the Triplet Selection Bias for Triplet Loss

Yu, Baosheng; Liu, Tongliang; Gong, Maoguo; Ding, Changxing; Tao, Dacheng

doi:10.1007/978-3-030-01231-1_5

Cited by 102 publications

(62 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The losses used in recently proposed deep metric learning methods [30,28,32,29,39,44] take into consideration of higher order relationships or global information and therefore achieve better performance. For example, Song et al [30] proposed a lifted structured loss function to consider all the positive and negative pairs within a batch.…”

Section: Related Workmentioning

confidence: 99%

Hardness-Aware Deep Metric Learning

Zheng

Chen

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

144

109

View full text Add to dashboard Cite

This paper presents a hardness-aware deep metric learning (HDML) framework. Most previous deep metric learning methods employ the hard negative mining strategy to alleviate the lack of informative samples for training. However, this mining strategy only utilizes a subset of training data, which may not be enough to characterize the global geometry of the embedding space comprehensively. To address this problem, we perform linear interpolation on embeddings to adaptively manipulate their hard levels and generate corresponding label-preserving synthetics for recycled training, so that information buried in all samples can be fully exploited and the metric is always challenged with proper difficulty. Our method achieves very competitive performance on the widely used CUB-200-2011, Cars196, and Stanford Online Products datasets.

show abstract

Section: Related Workmentioning

confidence: 99%

Hardness-Aware Deep Metric Learning

Zheng

Chen

et al. 2019

2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

144

109

View full text Add to dashboard Cite

show abstract

“…introduce synthetic data to reduce the negative effects of imbalance in datasets [22]. Yu et al propose an adaptive triplet selection for correcting the distribution shift and model bias [47]. Robinson et al show that the performance gaps in FR for various races can be reduced by adapting the decision thresholds for each race [36].…”

Section: Related Workmentioning

confidence: 99%

How (Not) to Measure Bias in Face Recognition Networks

Glüge

Amirian

Flumini

et al. 2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Within the last years Face Recognition (FR) systems have achieved human-like (or better) performance, leading to extensive deployment in large-scale practical settings. Yet, especially for sensible domains such as FR we expect algorithms to work equally well for everyone, regardless of somebody's age, gender, skin colour and/or origin. In this paper, we investigate a methodology to quantify the amount of bias in a trained Convolutional Neural Network (CNN) model for FR that is not only intuitively appealing, but also has already been used in the literature to argue for certain debiasing methods. It works by measuring the "blindness" of the model towards certain face characteristics in the embeddings of faces based on internal cluster validation measures. We conduct experiments on three openly available FR models to determine their bias regarding race, gender and age, and validate the computed scores by comparing their predictions against the actual drop in face recognition performance for minority cases. Interestingly, we could not link a crisp clustering in the embedding space to a strong bias in recognition rates-it is rather the opposite. We therefore offer arguments for the reasons behind this observation and argue for the need of a less naïve clustering approach to develop a working measure for bias in FR models.

show abstract

“…Successively, due to the exponential number of triplets we can generate, we adopt a simple and practical strategy to sample a subset of such triplets. We remind that triplet selection is an hard task and some research works are investigating how to smartly sample useful and informative subsets of triplet constraints [26]. It is out of the scope of this work supplying a method that competes with such strategies.…”

Section: Triplet Semi-supervised Deep Embedding Clusteringmentioning

confidence: 99%

“…On the other hand, we set up an easy and ready to use approach that well fits our scenario. Once the set of triplet constraints are chosen, we inject such background information into a deep-learning based clustering algorithm [26,9,27].…”

Section: Triplet Semi-supervised Deep Embedding Clusteringmentioning

confidence: 99%

“…where T is the set of triplet constraints, [x] + = max(0, x) is the hinge loss, ||x|| 2 2 is the squared L 2 norm of x, enc(x, Θ 1 ) is the encoder network, with weights parameters Θ 1 , applied on an example x, norm L2 is a function that performs the L 2 normalization of the output of the encoder and α is the margin hyperparameter usually involved in distance-based loss function to stretch the representation space [26]. We consider α equal to 1.0 since distances are derived by L 2 normalization.…”

Section: Triplet Semi-supervised Deep Embedding Clusteringmentioning

confidence: 99%

See 1 more Smart Citation

Deep Triplet-Driven Semi-supervised Embedding Clustering

Ienco

Pensa

2019

Discovery Science

View full text Add to dashboard Cite

In most real world scenarios, experts dispose of limited background knowledge that they can exploit for guiding the analysis process. In this context, semi-supervised clustering can be employed to leverage such knowledge and enable the discovery of clusters that meet the analysts' expectations. To this end, we propose a semi-supervised deep embedding clustering algorithm that exploits triplet constraints as background knowledge within the whole learning process. The latter consists in a two-stage approach where, initially, a low-dimensional data embedding is computed and, successively, cluster assignment is refined via the introduction of an auxiliary target distribution. Our algorithm is evaluated on real-world benchmarks in comparison with state-of-the-art unsupervised and semi-supervised clustering methods. Experimental results highlight the quality of the proposed framework as well as the added value of the new learnt data representation.

show abstract

Correcting the Triplet Selection Bias for Triplet Loss

Cited by 102 publications

References 38 publications

Hardness-Aware Deep Metric Learning

Hardness-Aware Deep Metric Learning

How (Not) to Measure Bias in Face Recognition Networks

Deep Triplet-Driven Semi-supervised Embedding Clustering

Contact Info

Product

Resources

About