Supervised two-stage transfer learning on imbalanced dataset for sport classification

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

2 Citations (Scopus)

Abstract

Sport classification is a crucial step for content analysis in a sport stream monitoring system. Training a reliable sport classifier can be a challenging task when the data is limited in amount and highly imbalanced. In this paper, we introduce a supervised two-stage transfer learning (Two-Stage-TL) method to solve the data shortage problem. It can progressively transfer features from a source domain to the target domain using a properly selected bridge domain. For the class imbalance issue, we compare several existing methods and demonstrate that the log-smoothing class weight is the most applicable way for this specific problem. Extensive experiments are conducted using ResNet50, VGG16, and Inception-ResNet-v2. The results show that Two-Stage-TL outperforms classical One-Stage-TL and achieves the best performance using log-smoothing class weight. The in-depth analysis is useful for researchers and developers in solving similar problems.

Original languageEnglish
Title of host publicationImage Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings
EditorsElisa Ricci, Nicu Sebe, Samuel Rota Bulò, Cees Snoek, Oswald Lanz, Stefano Messelodi
Place of PublicationCham
PublisherSpringer
Pages356-366
Number of pages11
ISBN (Electronic)978-3-030-30642-7
ISBN (Print)978-3-030-30641-0
DOIs
Publication statusPublished - 1 Jan 2019
Event20th International Conference on Image Analysis and Processing, ICIAP 2019 - Trento, Italy
Duration: 9 Sept 201913 Sept 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11751 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference20th International Conference on Image Analysis and Processing, ICIAP 2019
Country/TerritoryItaly
CityTrento
Period9/09/1913/09/19

Keywords

  • Class imbalance learning
  • Multimedia content analysis
  • Sport classification
  • Transfer learning

Fingerprint

Dive into the research topics of 'Supervised two-stage transfer learning on imbalanced dataset for sport classification'. Together they form a unique fingerprint.

Cite this