Accelerating the training of artificial neural networks

We develop software for the creation of synthetic training data for artificial neural networks (ANNs) from three-dimensional models. This enables the input data for ANNs to be generated at a much lower cost and with reduced risk. This makes the ANN-based evaluation of measurement data affordable even for SMEs.

Synthetic ANN training data for the road environment

© Fraunhofer IPM
© Fraunhofer IPM
Artificially created training dataset of an urban scene consisting of camera image and classification.

Artificial neural networks (ANNs) on the rise

Large amounts of highly complex measurement data are increasingly evaluated automatically using artificial neural networks (ANNs). Currently, ANNs are usually trained manually in an inefficient, error-prone process. Synthetic training data make ANN training more efficient and reliable. Fraunhofer IPM has proven how this can be done using software for interpreting road environment images. In a fully automated process, the software creates road scenes and converts them into both measurement and training data. Training based on this type of data guarantees reliability and significantly increased efficiency.

ANN training – still an expensive undertaking at present

Prompted by the need for automated environment analysis in autonomous driving, science and industry have shifted their focus toward the use of ANNs as a method of artificial intelligence (AI). The road environment is one of the most complex environments for an automated analysis, with significant regional variation. Unique training data must be created for every regional context for ANN training. To reliably train an ANN, several thousand images per object class (e.g. car, road, sidewalk or vegetation) are required on average for each region. To generate the training data required, mobile survey vehicles start by capturing images in the relevant regions. These are then manually annotated with the appropriate object classes. The annotation or classification of each image takes 20 to 40 minutes on average – valuable working time. Finally, the ANN is trained and evaluated. After the evaluation, follow-up training is usually necessary, for which further measurement data must be generated and classified.

Synthetic training data reduce costs dramatically

If it is possible to quickly and efficiently create synthetic training data, time-consuming field measurements will become redundant. There is no need for manual annotation and the findings from the assessment of the ANN can be used to optimize the synthetic generation of training data. All these processes are carried out on the computer, significantly reducing the training costs. This method also makes it possible to test the identifiability of new object classes without a high financial risk and therefore to quickly and easily assess the potential a given ANN offers for a specific application.