Workshop on "Generating synthetic image data for AI"
at the KI 2022 (virtual, hosted in Trier/Germany)

September 19th, 2022, 9:00 - 12:00

Our Workshop is co-located to the 45th German Conference on Artificial Intelligence (KI 2022), which will take place virtual (hosted in Trier/Germany) from September 19th to September 23rd, 2022. The link will follow as soon as possible.

While synthetic data generation is viable for most if not all machine learning application fields, this workshop focuses on both 2D and 3D Computer Vision. We aim to gather experts from both machine learning and simulation sides of the topic in order to explore current best practices, challenges and opportunities in using synthetic data.  

Topics of interest include the following and any related topics:

1. Physically accurate texture and shape modelling
2. Domain adaptation and randomization
3. Parameterized data generation techniques
4. Geometry vs. texture training impact
5. Simulation introduced data bias
6. Simulation prospects for industrial quality evaluation
7. Stress testing - using simulation to test solution robustness
8. Synthetic dataset evaluation

Diversity of perspectives is the key strength when it comes to establishing the field of simulation based machine learning. Therefore, we welcome experts coming from different fields such as stochastic geometry, computer graphics, mathematical modelling and machine learning as well as industry practitioners who have already started applying these methods in their solutions. 

AI, in particular deep learning, offers attractive solutions to challenging image processing tasks. Tedious development and parametrization of algorithmic solutions can be replaced by training a suitable convolutional neural network. 

Field-tested pre-trained architectures are at hand. Moreover, machine learning methods have higher potential to generalize than classical, tailor-suited solutions. However, ML approaches depend on the availability of a large amount of representative image data along with a ground truth, usually obtained by manual annotation.

There is a variety of image processing problems, where the required training data is not available: For example, for optical quality control in production, all defect types have to be trained. However, the more safety-critical a defect is, the less frequently it can be observed because it is avoided by all means during production. Manual annotation is costly and can lead to inconsistent results due to subjective decisions or simple operator fatigue. Consistent manual annotation is particularly challenging if not even impossible if it comes to segmentation of complex structures in higher dimensional image data like one pixel thick cracks in 3D images of concrete.

Synthetic images, generated based on realizations of stochastic geometry models offer an elegant way out. A wide variety of structure types can be generated. The within structure variation is naturally captured by the concept of random closed sets. The ground truth is for free.

Clearly, training based on synthetic images only raises new questions: How realistic do the structures and images have to be? What is the right way to measure this? When can a structure considered representative? Which characteristics of the synthetic training data are decisive for successful training?

Here you will find the schedule and downloads for our workshop:

Schedule 

Abstract BosnarPresentation Bosnar

Abstract Dahmen

Abstract FulirPresentation Fulir

Abstract Sauer Presentation Sauer

Abstract Schorr Presentation Schorr

If you have any questions, please contact us:

Niklas Rottmayer 

Sarah Miezal

back to the conference