Incorrect label detection using convolutional autoencoder in labeling task of optical motion capture data

Abstract

Although optical motion capture system has an advantage of accuracy compared to other motion capture methods, processing the data takes extra time and effort. For example, labeling for markers is necessary for optical motion capture. The labeling by the standard algorithm often fails, and then manual labeling by human must correct the errors. In this paper, we consider the method to modify mistaken label of markers by using convolutional autoencoder. We describe two kinds of strategies of methods. The first one is based on reproduction by using autoencoder. The second one uses the output of intermediate layer as feature. The neural network is trained with motion capture data and tested on the incorrect labeled data.