Time Series Data Augmentation Based on Motion Differences Between Trials

Abstract

This paper describes a method of data augmentation by focusing on the difference in motion of the time series data of different trials. The motion described here refers to the position and direction of objects. This compares a trial with a trial that is close to its movement and determines the amount of deformation. By considering not only the data to be augmented but also other data, we expect to generate more realistic data and improve learning accuracy. We believe that data augmentation to object motion will lead to application of data augmentation to robot motion such as manipulating objects.