This project was our first attempt at using synthetic media to train the computer vision model, YOLO. The MLAb is especially grateful to Mariana Fuentes, from the Marine Turtle Research, Ecology and Conservation Group, a professor in the Earth, Ocean and Atmospheric Science program at FSU, for her insights on turtle spotting, and turtles in general.
The project was a great success!
The research was presented at a Florida Undergraduate symposium, President’s Showcase. The paper went on to be published at an IEEE conference. The team used a custom Python library, written by John Sutor, that converts 3D files into a dataset of 2D images, labeled and ready to use as training data for a computer vision model. The team trained YOLO with synthetic data, which outperformed a model trained with authentic imagery. There should be no surprise there, as it is possible to control synthetic imagery, the variety of poses, and the “completeness” of the dataset.
