Training Convolutional Neural Networks with Synthesized Data for Object Recognition in Industrial Manufacturing

Visual tasks such as automated quality control or packaging require machines to be able to detect and identify objects automatically. In recent years object detection systems using deep learning have made significant advancements achieving better scores at a higher performance. However, these methods typically require large amounts of annotated images for training, which are costly and labor intensive to create. Therefore, it is an attractive alternative to generate the training data synthetically using computer-generated imagery (CGI). In this paper, we investigate how to add realistic texture to CAD objects to generate synthetic data for training of an instance segmentation network (Mask R-CNN) for recognition of manufacturing components. The results show that it is possible to create synthetic data with negligible human effort when using simple procedural materials.