Evaluating the Performance of Synthetic Visual Data for Real-Time Object Detection
Jonathan Adams, John Sutor, Ava Dodd, Erin Murphy · 2021
A method to produce synthetic visual data was developed for this study. The resulting synthetic imagery was used to train a YOLOv3 classification model to identify various species of sea turtles. The Blender library and Python were used to render and augment the images. Nine datasets were generated, each with different ratios of synthetic and authentic images. YOLOv3 was then trained and the performance resulting from each of the datasets was evaluated to measure which dataset produced the most confident model. The researchers optimized the classification model and demonstrated significant accuracy improvements, especially with models trained without authentic data. A dataset consisting of fully synthetic dataset performed best in testing, with an average confidence level of 9.68 percent greater than the fully authentic dataset.