The machine vision sector thrives on innovation. But how are new technologies created and how do they find their way into practice? Rebecca König, Research Engineer in the MVTec Software GmbH research department, gives us a tour of her workplace and explains the road from the initial idea to implementation.
It’s a wintry, Advent morning in December. We’re with Rebecca König in the kitchenette in MVTec’s head office in Munich. While the coffee flows into the mugs, we ask Rebecca what makes her job fulfilling. The researcher pauses briefly to think about it. She’s been working in MVTec’s research department since 2017, prior to which she studied mathematics at the Technical University of Munich.
On completing her degree, she wrote her master’s thesis at MVTec and worked here as student trainee. Whilst the milk froth tops the cappuccino, Rebecca gives us her answer: “The best part is realizing that the software algorithms you’ve developed don’t just work during simulations but also in practice.”
Customer feedback as a springboard for new developments
What attracts Rebecca to her work is also an important point for the machine vision sector as a whole. This is because the many minor and major innovations achieved by the company’s research and development team are what make the sector so dynamic. “Many of MVTec’s new developments result from specific customer problems. Here in the research department, it’s our task to find new approaches to things that can’t be resolved using existing methods,” explains Rebecca.
Datasets as a basis for new methods
The approach taken by Rebecca and her ten or so colleagues is based on university research. They start with literature-based research, checking whether scientific papers or even datasets already exist for another similar case. Such datasets are large image files that can be used to test new methods and technologies. The datasets play an important role. If no datasets yet exist, MVTec therefore develops its own ones and provides them to the research community.
In fact, it is even one of Rebecca and her colleagues’ aims to support the entire research community with their own research work in the form of datasets and publications. “We provide the datasets in downloadable format and regularly attend specialist conferences to present our work as well as to network and exchange ideas with other researchers. This results in exciting new ideas for our own work,” she says. However, the primary goal of MVTec’s research department is to implement the latest research results in the company’s software products as quickly as possible. To do this, the department not only has ten permanent staff members but also employs doctoral and master’s students.
A change of scene. We’ve now arrived at one of Rebecca’s several workstations, the MVTec app laboratory, where we find several devices, camera systems, and even a robot. This is the place where applications are developed for trade fairs, practical use cases are simulated, and the aforementioned datasets are created. Depending on the case, a dataset for developing new methods requires around 2000 images. These need to be created. “We face particularly high real-world requirements, making a comprehensive dataset essential for developing robust methods,” explains Rebecca.
Creativity is required for new algorithms
We’ve moved to yet another area of the company and are now sitting at Rebecca’s office workstation. After creating the datasets, the actual work starts here. Our next coffee is also ready. “There was no way that the last practical inquiry we received could be resolved using existing methods. That means we need to create something new. Deep learning may help us,” says Rebecca hopefully. Deep learning has been the on-trend technology in the sector for some years now. Like other companies, MVTec is investing time and brainpower into the development of new deep learning approaches.
Its latest success is “Global Context Anomaly Detection”, an evolved version of anomaly detection that can understand logical content and identify errors relating to the entire image.
For example, it can detect bottle labels that have slipped or been incorrectly printed, or missing components, for instance on circuit boards. The technology in this form is a world first.
Back to the current problem. Rebecca is in the process of developing a new deep learning network architecture. The algorithm programmed to do this, on which the network is based, will be subsequently fed with the dataset. In other words, the deep learning model will be trained using some of the data. The rest of the data will then be used to test and evaluate how well the algorithm resolves the problem. “There is more than one way to achieve a goal; creativity is required. For example, you can creatively achieve a goal by skillfully combining two complementary approaches in a single algorithm. This makes use of the benefits of both algorithms. Another example is to tweak various aspects of an algorithm to fully optimize every last bit of its performance. And that’s exactly what I enjoy about my work,” says Rebecca, adding: “Our team has intentionally given ourselves a great deal of free rein so that we can simply try things. We even undertake high-risk projects despite not knowing whether they’ll work. Failure is allowed.”
Integration of research results into MVTec software products
But how does a new technology bridge the gap from theory to practice? Once the new method is ready for the market, it’s about implementing it in one of MVTec’s software products. “To do this, we switch from the Research department to Development. We become part of the development team and help integrate the algorithm into the software,” says Rebecca. The new technology is then accessible to customers upon the next software release – and all that is left to do is await their feedback.