Back-dropout Transfer Learning for Action Recognition

Research output: Contribution to journalReview articleResearchpeer-review

Abstract

Transfer learning aims at adapting a model learned from source dataset to target dataset. It is a beneficial approach especially when annotating on the target dataset is expensive or infeasible. Transfer learning has demonstrated its powerful learning capabilities in various vision tasks. Despite transfer learning being a promising approach, it is still an open question how to adapt the model learned from the source dataset to the target dataset. One big challenge is to prevent the impact of category bias on classification performance. Dataset bias exists when two images from the same category, but from different datasets, are not classified as the same. To address this problem, we propose a transfer learning algorithm which takes advantage of misclassified images and applies back-dropout strategy to punish errors. We call this Negative Back-dropout Transfer Learning (NB-TL). Experimental results demonstrate the effectiveness of our algorithm. Moreover, our NB-TL algorithm has obtained 88.9% accuracy on UCF101 Action Recognition dataset, which achieves state-of-the-art performance.
Close

Details

Transfer learning aims at adapting a model learned from source dataset to target dataset. It is a beneficial approach especially when annotating on the target dataset is expensive or infeasible. Transfer learning has demonstrated its powerful learning capabilities in various vision tasks. Despite transfer learning being a promising approach, it is still an open question how to adapt the model learned from the source dataset to the target dataset. One big challenge is to prevent the impact of category bias on classification performance. Dataset bias exists when two images from the same category, but from different datasets, are not classified as the same. To address this problem, we propose a transfer learning algorithm which takes advantage of misclassified images and applies back-dropout strategy to punish errors. We call this Negative Back-dropout Transfer Learning (NB-TL). Experimental results demonstrate the effectiveness of our algorithm. Moreover, our NB-TL algorithm has obtained 88.9% accuracy on UCF101 Action Recognition dataset, which achieves state-of-the-art performance.
Original languageEnglish
JournalIET Computer Vision
Volume12
Issue number4
Pages (from-to)1-15
Number of pages15
ISSN1751-9632
DOI
Publication statusPublished - 2018
Publication categoryResearch
Peer-reviewedYes
ID: 240061350