Synthetic training data

The quality of training data is the key factor influencing the performance of an AI. However, data generation is also the main cost driver. Both the selection of suitable raw data and manual annotation are very time-consuming.

However, the effort and costs can be significantly reduced. For many use cases, measurement scenarios can be parametrically generated using freely available models, which then produce synthetic measurement data by simulating the measurement process. This could be a photorealistic computer graphic (“rendering”) for an image-based AI or a point cloud from a simulated laser scanner. The key advantage is that: Since the scene is composed of known models to which a class can be assigned, this information can be transferred to the measurement data, allowing the synthetic data to be used as training data without further processing. Fraunhofer IPM develops software for automatically generating synthetic training data from three-dimensional models. When rendering the respective scene, we draw on our extensive expertise in modelling and measurement technology.

Synthetic data can never perfectly replicate reality. There remains what is known as the "domain gap." The neural network must learn to transfer the features it has learned from synthetic data to real data. We address this issue by using a combination of synthetic and real training data. Second, we optimize the generation of training data, not to produce the most detailed and realistic scenes possible, but to focus on parameters relevant to the network’s learning success (e.g., correctly simulating measurement noise in a point cloud).