Training Deep Neural Networks for Detecting Drinking Glasses Using Synthetic Images

This study presents an approach of using synthetically rendered images for training deep neural networks on object detection. A new plug-in for the computer graphics modelling software Blender was developed that can generate large numbers of photo-realistic ray-traced images and include meta information as training labels. The performance of the deep neural network DetectNet is evaluated using training data comprising synthetically rendered images and digital photos of drinking glasses. The detection accuracy is determined by comparing bounding boxes using intersection over union technique. The detection experiments using real-world and synthetic image data resulted in comparable results and the performance increased when using a pre-trained GoogLeNet model. The experiments demonstrated that training deep neural networks for object detection on synthetic data is effective and the proposed approach can be useful for generating large labelled image data sets to enhance the performance of deep neural networks on specific object detection tasks.