Stochastic Class-Based Hard Example Mining for Deep Metric Learning

Suh, Young-Ger; Han, Bohyung; Kim, Wonsik; Lee, Kyoung Mu

doi:10.1109/cvpr.2019.00742

Cited by 123 publications

(56 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…[3], and DREML [5]) with a large margin. Impressively, even the single-head version of HDhE (Sh-HDhE) having only the pruned head (Last layer is Res4_1) shows a competitive performance to the latest algorithms (SCHM [8], SoftTriple [9], and Proxy-Anchor [10]). Furthermore, the proposed HDhE gives a performance improvement of over 4% of Recall@1 (CUB-200) from the single-head version of HDhE (Sh-HDhE).…”

Section: E Comparison To State-of-the-art Methodsmentioning

confidence: 99%

“…Each sub-layer learns in a successive manner, in which the previous layer learns the samples and the next layer learns the re-weighed sample through the gradient generated from the previous layer. However, the algorithms [4], [8], which appeared later, can achieve higher performance without ensemble than the above algorithms, while using the same backbone network (inception-v1).…”

Section: Related Work a Ensemble For The Deep Metric Learningmentioning

confidence: 98%

“…Hence this one is not recommended in order to use feature vectors due to excessively large dimensions, even though it shows the best performance, as seen in Table 1. • The variant "adding linear layer" variant is a method commonly used in other algorithms [3], [4], [8]. However, this method has a large performance drop compared to the original one, as shown in the second row of Table 1.…”

Section: A Dimension-reduced Feature Vectormentioning

confidence: 99%

See 2 more Smart Citations

Heterogeneous Double-Head Ensemble for Deep Metric Learning

Ro¹,

Choi²

2020

IEEE Access

View full text Add to dashboard Cite

The structure of a multi-head ensemble has been employed by many algorithms in various applications including deep metric learning. However, their structures have been empirically designed in a simple way such as using the same head structure, which leads to a limited ensemble effect due to lack of head diversity. In this paper, for an elaborate design of the multi-head ensemble structure, we establish design concepts based on three structural factors: designing the feature layer for extracting the ensemblefavorable feature vector, designing the shared part for memory savings, and designing the diverse multiheads for performance improvement. Through rigorous evaluation of variants on the basis of the design concepts, we propose a heterogeneous double-head ensemble structure that drastically increases ensemble gain along with memory savings. In verifying experiments on image retrieval datasets, the proposed ensemble structure outperforms the state-of-the-art algorithms by margins of over 5.3%, 6.1%, 5.9%, and 1.8% in CUB-200, Car-196, SOP, and Inshop, respectively. INDEX TERMS Ensemble learning, multi-head structure, deep metric learning, deep architecture design, image retrieval

show abstract

Section: E Comparison To State-of-the-art Methodsmentioning

confidence: 99%

Section: Related Work a Ensemble For The Deep Metric Learningmentioning

confidence: 98%

Section: A Dimension-reduced Feature Vectormentioning

confidence: 99%

See 1 more Smart Citation

Heterogeneous Double-Head Ensemble for Deep Metric Learning

Ro¹,

Choi²

2020

IEEE Access

View full text Add to dashboard Cite

show abstract

“…Consequently, two samples from the same subspace have smaller distances than two from different clusters, which can be interpreted as a proxy to the mining of meaningful relationships. Stochastic class-based hard example mining [52] is proposed to mine hard examples effectively.…”

Section: Mining and Weighting Non-trivial Examplesmentioning

confidence: 99%

Ranked List Loss for Deep Metric Learning

Wang

Yang

Kodirov³

et al. 2021

IEEE Trans. Pattern Anal. Mach. Intell.

View full text Add to dashboard Cite

The objective of deep metric learning (DML) is to learn embeddings that can capture semantic similarity and dissimilarity information among data points. Existing pairwise or tripletwise loss functions used in DML are known to suffer from slow convergence due to a large proportion of trivial pairs or triplets as the model improves. To improve this, ranking-motivated structured losses are proposed recently to incorporate multiple examples and exploit the structured information among them. They converge faster and achieve state-of-the-art performance. In this work, we unveil two limitations of existing ranking-motivated structured losses and propose a novel ranked list loss to solve both of them. First, given a query, only a fraction of data points is incorporated to build the similarity structure. Consequently, some useful examples are ignored and the structure is less informative. To address this, we propose to build a set-based similarity structure by exploiting all instances in the gallery. The learning setting can be interpreted as few-shot retrieval: given a mini-batch, every example is iteratively used as a query, and the rest ones compose the galley to search, i.e., the support set in few-shot setting. The rest examples are split into a positive set and a negative set. For every mini-batch, the learning objective of ranked list loss is to make the query closer to the positive set than to the negative set by a margin. Second, previous methods aim to pull positive pairs as close as possible in the embedding space. As a result, the intraclass data distribution tends to be extremely compressed. In contrast, we propose to learn a hypersphere for each class in order to preserve useful similarity structure inside it, which functions as regularisation. Extensive experiments demonstrate the superiority of our proposal by comparing with the state-of-the-art methods on the fine-grained image retrieval task. Our source code is available online: https://github.com/XinshaoAmos Wang/Ranked-List-Loss-for-DML.

show abstract

“…Deep metric learning aims to project images to a low dimensional embedding space, in which the images with similar semantics are clustered together [27,23,4]. The most popular paradigm is to employ the triplet loss to penalize the positive pair or negative pair or both of them within a triplet [25].…”

Section: Metric Learningmentioning

confidence: 99%

Instance Hard Triplet Loss for In-video Person Re-identification

et al. 2020

View full text Add to dashboard Cite

Traditional Person Re-identification (ReID) methods mainly focus on cross-camera scenarios, while identifying a person in the same video/camera from adjacent subsequent frames is also an important question, for example, in human tracking and pose tracking. We try to address this unexplored in-video ReID problem with a new large-scale video-based ReID dataset called PoseTrack-ReID with full images available and a new network structure called ReID-Head, which can extract multi-person features efficiently in real time and can be integrated with both one-stage and two-stage human or pose detectors. A new loss function is also required to solve this new in-video problem. Hence, a triplet-based loss function with an online hard example mining designed to distinguish persons in the same video/group is proposed, called instance hard triplet loss, which can be applied in both cross-camera ReID and in-video ReID. Compared with the widely-used batch hard triplet loss, our proposed loss achieves competitive performance and saves more than 30% of the training time. We also propose an automatic reciprocal identity association method, so we can train our model in an unsupervised way, which further extends the potential applications of in-video ReID. The PoseTrack-ReID dataset and code will be publicly released.

show abstract

Stochastic Class-Based Hard Example Mining for Deep Metric Learning

Cited by 123 publications

References 24 publications

Heterogeneous Double-Head Ensemble for Deep Metric Learning

Heterogeneous Double-Head Ensemble for Deep Metric Learning

Ranked List Loss for Deep Metric Learning

Instance Hard Triplet Loss for In-video Person Re-identification

Contact Info

Product

Resources

About