Relationship-Aware Hard Negative Generation in Deep Metric Learning

Jiaqi HUANG, Yong FENG*, Mingliang ZHOU, Baohua QIANG

*Corresponding author for this work

Research output: Book Chapters | Papers in Conference ProceedingsConference paper (refereed)Researchpeer-review


Data relationships and the impact of synthetic loss have not been concerned by previous sample generation methods, which lead to bias in model training. To address above problem, in this paper, we propose a relationship-aware hard negative generation (RHNG) method. First, we build a global minimum spanning tree for all categories to measure the data distribution, which is used to constrain hard sample generation. Second, we construct a dynamic weight parameter which reflects the convergence of the model to guide the synthetic loss to train the model. Experimental results show that the proposed method outperforms the state-of-the-art methods in terms of retrieval and clustering tasks.

Original languageEnglish
Title of host publicationKnowledge Science, Engineering and Management : 13th International Conference, KSEM 2020, Proceedings
EditorsGang LI, Heng Tao SHEN, Ye YUAN, Xiaoyang WANG, Huawen LIU, Xiang ZHAO
PublisherSpringer, Cham
Number of pages13
ISBN (Electronic)9783030553937
ISBN (Print)9783030553920
Publication statusPublished - 2020
Externally publishedYes
Event13th International Conference on Knowledge Science, Engineering and Management, KSEM 2020 - Hangzhou, China
Duration: 28 Aug 202030 Aug 2020

Publication series

NameLecture Notes in Computer Science
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference13th International Conference on Knowledge Science, Engineering and Management, KSEM 2020

Bibliographical note

Publisher Copyright:
© 2020, Springer Nature Switzerland AG.


  • Deep metric learning
  • Distribution quantification
  • Minimum spanning tree
  • Relationship preserving
  • Sample generation


Dive into the research topics of 'Relationship-Aware Hard Negative Generation in Deep Metric Learning'. Together they form a unique fingerprint.

Cite this