WeakCounter: Acceleration-based Repetition Counting of Actions with Weakly Supervised Learning

Abstract

This study presents WeakCounter, which is a weakly supervised method for the repetition counting of a human action using a body-worn inertial sensor. WeakCounter is composed of two novel components: i) an attention-based network that can be trained on a weak label, which is defined to specify only the number of repetitions of an action included in an input data segment in this study, and ii) label diversification that enhances training of the network.