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

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

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 Sep 201913 Sep 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
CountryItaly
CityTrento
Period9/09/1913/09/19

Fingerprint

Transfer Learning
Sports
Smoothing
Content Analysis
Shortage
Monitoring System
Classifiers
Classifier
Target
Monitoring
Demonstrate
Experiment
Class
Experiments

Keywords

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

Cite this

Bi, T., Jarnikov, D., & Lukkien, J. (2019). Supervised two-stage transfer learning on imbalanced dataset for sport classification. In E. Ricci, N. Sebe, S. Rota Bulò, C. Snoek, O. Lanz, & S. Messelodi (Eds.), Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings (pp. 356-366). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11751 LNCS). Cham: Springer. https://doi.org/10.1007/978-3-030-30642-7_32
Bi, Tianyu ; Jarnikov, Dmitri ; Lukkien, Johan. / Supervised two-stage transfer learning on imbalanced dataset for sport classification. Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings. editor / Elisa Ricci ; Nicu Sebe ; Samuel Rota Bulò ; Cees Snoek ; Oswald Lanz ; Stefano Messelodi. Cham : Springer, 2019. pp. 356-366 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{b48f544d142a44898cefa53cb2d99cd9,
title = "Supervised two-stage transfer learning on imbalanced dataset for sport classification",
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.",
keywords = "Class imbalance learning, Multimedia content analysis, Sport classification, Transfer learning",
author = "Tianyu Bi and Dmitri Jarnikov and Johan Lukkien",
year = "2019",
month = "1",
day = "1",
doi = "10.1007/978-3-030-30642-7_32",
language = "English",
isbn = "978-3-030-30641-0",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer",
pages = "356--366",
editor = "Elisa Ricci and Nicu Sebe and {Rota Bul{\`o}}, Samuel and Cees Snoek and Oswald Lanz and Stefano Messelodi",
booktitle = "Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings",
address = "Germany",

}

Bi, T, Jarnikov, D & Lukkien, J 2019, Supervised two-stage transfer learning on imbalanced dataset for sport classification. in E Ricci, N Sebe, S Rota Bulò, C Snoek, O Lanz & S Messelodi (eds), Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11751 LNCS, Springer, Cham, pp. 356-366, 20th International Conference on Image Analysis and Processing, ICIAP 2019, Trento, Italy, 9/09/19. https://doi.org/10.1007/978-3-030-30642-7_32

Supervised two-stage transfer learning on imbalanced dataset for sport classification. / Bi, Tianyu; Jarnikov, Dmitri; Lukkien, Johan.

Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings. ed. / Elisa Ricci; Nicu Sebe; Samuel Rota Bulò; Cees Snoek; Oswald Lanz; Stefano Messelodi. Cham : Springer, 2019. p. 356-366 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11751 LNCS).

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

TY - GEN

T1 - Supervised two-stage transfer learning on imbalanced dataset for sport classification

AU - Bi, Tianyu

AU - Jarnikov, Dmitri

AU - Lukkien, Johan

PY - 2019/1/1

Y1 - 2019/1/1

N2 - 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.

AB - 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.

KW - Class imbalance learning

KW - Multimedia content analysis

KW - Sport classification

KW - Transfer learning

UR - http://www.scopus.com/inward/record.url?scp=85072983920&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-30642-7_32

DO - 10.1007/978-3-030-30642-7_32

M3 - Conference contribution

AN - SCOPUS:85072983920

SN - 978-3-030-30641-0

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 356

EP - 366

BT - Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings

A2 - Ricci, Elisa

A2 - Sebe, Nicu

A2 - Rota Bulò, Samuel

A2 - Snoek, Cees

A2 - Lanz, Oswald

A2 - Messelodi, Stefano

PB - Springer

CY - Cham

ER -

Bi T, Jarnikov D, Lukkien J. Supervised two-stage transfer learning on imbalanced dataset for sport classification. In Ricci E, Sebe N, Rota Bulò S, Snoek C, Lanz O, Messelodi S, editors, Image Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings. Cham: Springer. 2019. p. 356-366. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-030-30642-7_32