Few-Shot and Weakly Supervised Repetition Counting With Body-Worn Accelerometers

Nishino, Yuuki and Maekawa, Takuya and Hara, Takahiro (2022) Few-Shot and Weakly Supervised Repetition Counting With Body-Worn Accelerometers. Frontiers in Computer Science, 4. ISSN 2624-9898

[thumbnail of pubmed-zip/versions/1/package-entries/fcomp-04-925108/fcomp-04-925108.pdf] Text
pubmed-zip/versions/1/package-entries/fcomp-04-925108/fcomp-04-925108.pdf - Published Version

Download (2MB)

Abstract

This study investigates few-shot weakly supervised repetition counting of a human action such as workout using a wearable inertial sensor. We present WeakCounterF that leverages few weakly labeled segments containing occurrences of a target action from a target user to achieve precise repetition counting. Here, a weak label is defined to specify only the number of repetitions of an action included in an input data segment in this study, facilitating preparation of datasets for repetition counting. First, WeakCounterF leverages data augmentation and label diversification techniques to generate augmented diverse training data from weakly labeled data from users other than a target user, i.e., source users. Then, WeakCounterF generates diverse weakly labeled training data from few weakly labeled training data from the target user. Finally, WeakCounterF trains its repetition counting model composed of an attention mechanism on the augmented diversified data from the source users, and then fine-tunes the model on the diversified data from the target user.

Item Type: Article
Subjects: Science Repository > Computer Science
Depositing User: Managing Editor
Date Deposited: 24 Dec 2022 07:16
Last Modified: 27 Dec 2023 05:52
URI: http://research.manuscritpub.com/id/eprint/549

Actions (login required)

View Item
View Item