Leveraging Virtual Environments To Train A Deep Learning Algorithm

Keywords

Deep learning; Machine learning; Photogrammetry; Virtual environments

Abstract

Open source datasets are the typical source used to train computers to accurately detect visual objects (e.g., humans, animals, and inanimate objects) through various machine learning methods (e.g., Deep Learning (DL)). This data, however, is not feasible for use in the military domain. In this paper, a comparative analysis of real and virtual training data is provided, using the You Only Look Once (YOLO) Convolutional Neural Network (CNN) model. The main concern of this paper is to verify the process and accuracy of using a domain-specific U.S. Army Virtual Environment (VE), in contrast to a Real Environment (RE) dataset, with DL. Comparative results suggest that substituting a VE to provide training data for the DL model saves manual labour while maintaining a quality precision-recall curve.

Publication Date

1-1-2018

Publication Title

17th International Conference on Modeling and Applied Simulation, MAS 2018

Number of Pages

48-54

Document Type

Article; Proceedings Paper

Personal Identifier

scopus

Socpus ID

85056631010 (Scopus)

Source API URL

https://api.elsevier.com/content/abstract/scopus_id/85056631010

This document is currently not available here.

Share

COinS