Working on the project for creating a system for industrial defects detection with the use of drone shooting and the results analysis by a neural network, we faced a lack of materials for the dataset. We decided to create a tool for generation of artificial marked images (in the form of bitmasks and bounding boxes), and the final result would be used to enrich the dataset for training deep convolutional architectures. With the help of the synthetic data generator we managed to save much time and money comparing to the manual dataset collection and creation.
Read more:
Article. Synthetic data generator. Neural network training for industrial defectoscopy