Synthetic Ground Truth Generation for Object Recognition Evaluation: A Scalable System for Parameterized Creation of Annotated Images

Abstract

The number of application areas for object recognition are on the rise. Usable data are often limited or not well prepared to adapt to research. For image-based recognition, extensive training data is required in order to achieve precise object recognition with good repeatability. In order to generate training data with a high variance of individual parameters required for indoor localization, we developed a pipeline for Synthetic ground truth generation. This pipeline can be used to generate specific training data for object recognition in large amounts. Another field of application is the testing of the behavior of trained networks against specific parameters.