Weakly-supervised action recognition and localization via knowledge transfer

Haichao Shi, Xiaoyu Zhang*, Changsheng Li

*Corresponding author for this work

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

1 Citation (Scopus)

Abstract

Action recognition and localization has attracted much attention in the past decade. However, a challenging problem is that it typically requires large-scale temporal annotations of action instances for training models in untrimmed video scenarios, which is not practical in many real-world applications. To alleviate the problem, we propose a novel weakly-supervised action recognition framework for untrimmed videos to use only video-level annotations to transfer information from publicly available trimmed videos to assist in model learning, namely KTUntrimmedNet. A two-stage method is designed to guarantee an effective transfer strategy: Firstly, the trimmed and untrimmed videos are clustered to find similar classes between them, so as to avoid negative information transfer from trimmed data. Secondly, we design an invariant module to find common features between trimmed videos and untrimmed videos for improving the performance. Extensive experiments on the standard benchmark datasets, THUMOS14 and ActivityNet1.3, clearly demonstrate the efficacy of our proposed method when compared with the existing state-of-the-arts.

Original languageEnglish
Title of host publicationPattern Recognition and Computer Vision- 2nd Chinese Conference, PRCV 2019, Proceedings, Part I
EditorsZhouchen Lin, Liang Wang, Tieniu Tan, Jian Yang, Guangming Shi, Nanning Zheng, Xilin Chen, Yanning Zhang
PublisherSpringer
Pages205-216
Number of pages12
ISBN (Print)9783030316532
DOIs
Publication statusPublished - 2019
Externally publishedYes
Event2nd Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2019 - Xi'an, China
Duration: 8 Nov 201911 Nov 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11857 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference2nd Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2019
Country/TerritoryChina
CityXi'an
Period8/11/1911/11/19

Keywords

  • Action localization
  • Action recognition
  • Knowledge transfer

Fingerprint

Dive into the research topics of 'Weakly-supervised action recognition and localization via knowledge transfer'. Together they form a unique fingerprint.

Cite this