Synthetic image data generation for the training of a segmentation algorithm for machine tool recognition
Manuel Belke, Tim Peters, Oliver Petrović, Christian Brecher · 2025
Semantic understanding of the environment is crucial for indoor navigation. To train the machine learning algorithm, a large dataset is necessary, which can be generated synthetically to reduce manual work to label the data. A synthetic data generation pipeline is developed using Nvidia Isaac Sim to randomly generate various datasets for industrial environments. These environments contain machine tools, distractor objects and different lighting conditions. Six different randomization strategies are evaluated by training two different YOLO models with the synthetic data. The trained models are evaluated with synthetic and real data. The best Sim2Real transfer strategy in this scenario involves using background images of real industrial environments.