Increasing Oversampling Diversity for Long-Tailed Visual Recognition

Liuyu Xiang, Guiguang Ding*, Jungong Han

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference Proceeding (Non-Journal item)

2 Citations (SciVal)


The long-tailed data distribution in real-world greatly increases the difficulty of training deep neural networks. Oversampling minority classes is one of the commonly used techniques to tackle this problem. In this paper, we first analyze that the commonly used oversampling technique tends to distort the representation learning and harm the network’s generalizability. Then we propose two novel methods to increase the minority feature’s diversity to alleviate such issue. Specifically, from the data perspective, we propose a mixup-based Synthetic Minority Over-sampling TEchnique called mixSMOTE, where tail class samples are synthesized from head classes so that a balanced training distribution can be obtained. Then from the model perspective, we propose Gradient Re-weighting Module (GRM) to re-distribute each instance’s gradient contribution to the representation learning network. Extensive experiments on the long-tailed benchmark CIFAR10-LT, CIFAR100-LT and ImageNet-LT demonstrate the effectiveness of our proposed method.

Original languageEnglish
Title of host publicationArtificial Intelligence - 1st CAAI International Conference, CICAI 2021, Proceedings
EditorsLu Fang, Yiran Chen, Guangtao Zhai, Jane Wang, Ruiping Wang, Weisheng Dong
PublisherSpringer Nature
Number of pages12
ISBN (Electronic)9783030930462
ISBN (Print)9783030930455
Publication statusPublished - 01 Jan 2022
Event1st CAAI International Conference on Artificial Intelligence, CICAI 2021 - Hangzhou, China
Duration: 05 Jun 202106 Jun 2021

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13069 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference1st CAAI International Conference on Artificial Intelligence, CICAI 2021
Period05 Jun 202106 Jun 2021


  • Data imbalance
  • Long-tailed classification
  • Oversampling


Dive into the research topics of 'Increasing Oversampling Diversity for Long-Tailed Visual Recognition'. Together they form a unique fingerprint.

Cite this