Synthetic Dataset Generation and Adaptation for Human Detection

Abstract

In this report, we introduce an effort to use synthetic imagery to test machine learning classifiers as an alternative to large-scale unmanned aerial vehicle based data collections. The benefits of synthetic data are that it allows for the generation of a highly diverse dataset, gives you the ability to finely control experiments, and provides automatic annotation. We have demonstrated that synthetic data can be generated using a game engine and then used to perform experiments to validate machine learning classifiers. In these experiments, we visualized the expected performance of three classifiers Tiny-YOLO, YOLOv3, and RetinaNetand compared the performance of the classifiers to one another.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Nov 13, 2020
Accession Number
AD1115446

Entities

People

  • Damon Conover
  • Eung Joo Lee
  • Heesung Kwon
  • Jie Yan

Organizations

  • United States Army Research Laboratory

Tags

Communities of Interest

  • Autonomy
  • Counter WMD
  • Energy and Power Technologies

DTIC Thesaurus Topics

  • Accuracy
  • Aircrafts
  • Deep Learning
  • Demographic Cohorts
  • Department Of Defense
  • Detection
  • Information Science
  • Learning
  • Machine Learning
  • Materials
  • Standards
  • Three Dimensional
  • Training
  • Unmanned
  • Unmanned Aerial Vehicles
  • Vehicles
  • Weapons Of Mass Destruction

Fields of Study

  • Computer science

Readers

  • Image Processing and Computer Vision.
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks
  • Autonomy