Learning for Free: Object Detectors Trained on Synthetic Data
Publication Date
1-4-2021
Document Type
Conference Proceeding
Publication Title
Proceedings of the 2021 15th International Conference on Ubiquitous Information Management and Communication, IMCOM 2021
DOI
10.1109/IMCOM51814.2021.9377353
Abstract
A picture is worth a thousand words, or if you want it labeled, it's worth about ten cents per bounding box. Data is the fuel that powers modern technologies run by AI engines. High quality data is important to produce accurate machine learning models. Acquiring high quality labeled data however, can be expensive and time consuming. For small companies, academic researchers, or hobbyists, gathering large datasets that are not already publicly available is challenging. This research paper will describe the ability to generate labeled image data synthetically which can be used in supervised learning for object detection. This paper describes a system using 3D modeling software in conjunction with Generative Adversarial Networks and image augmentation that can create a diverse dataset of images containing objects with bounding boxes and labels. The result of this effort is an accurate object detector in an environment of aerial surveillance with no cost to the end user.
Department
Computer Science
Recommended Citation
Charles Thane MacKay and Teng Sheng Moh. "Learning for Free: Object Detectors Trained on Synthetic Data" Proceedings of the 2021 15th International Conference on Ubiquitous Information Management and Communication, IMCOM 2021 (2021). https://doi.org/10.1109/IMCOM51814.2021.9377353
