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

Onderzoeksoutput: Hoofdstuk in Boek/Rapport/CongresprocedureConferentiebijdrageAcademicpeer review

1 Citaat (Scopus)

Samenvatting

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.

Originele taal-2Engels
TitelImage Analysis and Processing – ICIAP 2019 - 20th International Conference, Proceedings
RedacteurenElisa Ricci, Nicu Sebe, Samuel Rota Bulò, Cees Snoek, Oswald Lanz, Stefano Messelodi
Plaats van productieCham
UitgeverijSpringer
Pagina's356-366
Aantal pagina's11
ISBN van elektronische versie978-3-030-30642-7
ISBN van geprinte versie978-3-030-30641-0
DOI's
StatusGepubliceerd - 1 jan 2019
Evenement20th International Conference on Image Analysis and Processing, ICIAP 2019 - Trento, Italië
Duur: 9 sep 201913 sep 2019

Publicatie series

NaamLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11751 LNCS
ISSN van geprinte versie0302-9743
ISSN van elektronische versie1611-3349

Congres

Congres20th International Conference on Image Analysis and Processing, ICIAP 2019
Land/RegioItalië
StadTrento
Periode9/09/1913/09/19

Vingerafdruk

Duik in de onderzoeksthema's van 'Supervised two-stage transfer learning on imbalanced dataset for sport classification'. Samen vormen ze een unieke vingerafdruk.

Citeer dit