SIMULATION OF IMAGE DATA TO SUPPORT THE TRAINING OF CONVOLUTIONAL NEURAL NETWORKS FOR OBJECTS RECOGNITION

Abstract

The recognition of logistics objects is an essential prerequisite for the optimization of operational logistics processes and can be performed among others via image-based methods. However, the lack of available data for training domain-specific recognition models remains a practical problem. For this reason, we present an approach to solving this problem. The core principle of our approach is the automated generation of image data from 3D models, in which the appearance of the objects varies through variations of different parameters. The first results are promising: Without any real image data, we have created a neural network for recognition of real objects with a recall quality of 86%.

This paper is a secondary publication of the article presented at the COMEC 2019 conference in Cuba, published with the permissions of the organization board of the conference and the authors.