Master of Science (MS)
sythetic data generation, image video data, YOLO, GAN
A picture is worth a thousand words, or if you want it labeled, it’s worth about four cents per bounding box. Data is the fuel that powers modern technologies run by artificial intelligence engines which is increasingly valuable in today’s industry. High quality labeled data is the most important factor in producing accurate machine learning models which can be used to make powerful predictions and identify patterns humans may not see. Acquiring high quality labeled data however, can be expensive and time consuming. For small companies, academic researchers, or machine learning hobbyists, gathering large datasets for a specific task that are not already publicly available is challenging. This research paper describes the techniques used to generate labeled image data synthetically which can be used in supervised learning for object detection. Technologies such as 3D modeling software in conjunction with Generative Adversarial Networks and image augmentation can create a realistic and diverse image dataset with bounding boxes and labels. The result of our effort is an accurate object detector in an environment of aerial surveillance with no cost to the end user. We achieved a best average precision score of 0.76 to classify and detect cars from an aerial perspective using a mix of GAN-refined data along with randomized synthetic data.
MacKay, Charles Thane, "Learning for Free – Object Detectors Trained on Synthetic Data" (2019). Master's Projects. 738.