Development of a Virtual Environment Based Image Generation Tool for Neural Network Training

Abstract

We present a computational tool to generate visual and descriptive data used as additional training images for neural networks involved in image recognition tasks. The work is inspired by the problem posed to acquire enough data, in order to train service robots, with the goal of improving the range of objects in the environment with which they can interact. The tool provides a framework that allows users to easily setup different environments with the visual information needed for the training, accordingly to their needs. The tool was developed with the Unity engine, and it was designed to be able to import external prefabs. These models are standardized and catalogued into lists, which are accessed to create more complex and diverse virtual environments. Another component of the tool adds an additional layer of complexity by creating randomized environments with different conditions (scale, position and orientation of objects, and environmental illumination). The performance of the created dataset was tested by training the information on the YOLO-V3 (You Only Look Once) architecture and testing on both artificial and real images.