Revolutionizing Deep Learning Algorithms Training with SIGEDA

In the world artificial intelligence and computer vision, one of the biggest challenges is having suitable datasets to train deep learning algorithms. The tool SIGEDAcreated at ITCL, is presented as an innovative solution to this problem, allowing the generation of synthetic datasets that facilitate the rendering of thousands of labeled images and videos.

Martijn van GasterenCoordinator of international scientific and innovation projects ITCLand scientific and technical coordinator of the European project HosmartAIhas over twenty years of experience in research and innovation management in European projects. He is passionate about contributing to the implementation of European Union policies and strives to “Make society a little better with technology”always collaborating with people who benefit from technology but don’t necessarily understand it.

He has worked on projects in many industries, including healthcare, transportation, education and industryHe has also evaluated projects for the European Commission and other European organisations that fund research and innovation.

The SIGEDA project was created within the framework of the European HosmartAI project a budget of 12 million euros and 10 million euros in subsidies from the European CommissionThe consortium includes 24 organisations such as companies, hospitals, universities, technology centres and associations from across Europe.

The main goal of HosmartAI was promote the use of artificial intelligence and robotics in hospitals through 8 major pilot projects and lasted 3 years and 4 months, ending in May last year.

Van Gasteren explains that photo- and image-based technologies are already fully integrated into our lives and are becoming increasingly important, from our smartphones to surveillance cameras and in industry. The algorithms used in so-called computer vision have advanced significantly with the advent of artificial intelligence, but they only work well if they are trained correctly.

This is where SIGEDA comes in – a great tool for generating large numbers of images that are used to train AI algorithms.

So, it is a technology for creating synthetic datasets that allow thousands of labeled images and/or videos to be presented to train artificial intelligence (deep learning) algorithms for artificial vision.

In many cases, there are very few suitable images or videos. This service allows AI developers to train their algorithms using images or videos created in virtual reality. For example, for human detection, a huge variety of “characters” can be created thanks to the ability to customize VR avatars, thereby avoiding bias and ethical issues. Moreover, since these images are created on a computer, the exact position of the person or object in the environment is known, and there is no need to estimate the position of the photo.

The development started 5 years ago in another European project called WorkingAge, which sought to evaluate the ergonomic posture of workers without uncomfortable sensors on their bodies. They chose to use a camera and artificial vision technology. The lack of specific photos of workers in a studio, office, and car factory gave them the idea to create synthetic images. Since then, they have been providing the service of creating image sets to clients: developers of computer vision solutions.

At HosmartAI, they have expanded the technology from images to video, and this adds a time dimension. This extends the detection of people from postures to movements, such as when people fall.

SIGEDA was used for machine vision with a humanoid robot in one of the pilot tests at a hospital in Slovenia. The robot helps patients with daily care and can, for example, detect patient falls and provide support during breathing exercises.

Together with the University Hospital of Brussels and the Free University of Brussels, they are also conducting research into how these technologies can be effectively used to model the growth of brain tumors. Imaging technologies are very important in many areas of healthcare, and SIGEDA is helping to make them more accurate, faster, and therefore cheaper.

In addition to applications in healthcare and working conditions, it can also be used in security or agriculture, for example, to monitor the growth of crops. Another example is monitoring industrial processes. Computer vision is used in all industries, and almost all of these applications use AI, and thousands of images are needed to train the AI.

Van Gasteren notes that a client who wants to train their AI saves a lot of time because SIGEDA automatically adds labels to images; labeling photos is an expensive manual job, and less accurate. This labeling is necessary to inform the AI ​​system about what is in the image. For example, in the case of human posture, the labels indicate the position of the main joints that determine posture: neck, shoulders, elbows, wrists, hips, knees and feet, in SIGEDA these parameters are predefined in each image.

In addition, in the case of humans, there are no privacy issues, since it is not real people that are used, but perfectly adaptable virtual entertainment. In this way, it is possible to create a representative distribution by gender, age, race, size, etc. of the population, avoiding important ethical issues.

The global synthetic data generation market is projected to grow from $300 million in 2023 to $2.1 billion in 2028; a very significant growth driven by concerns about privacy and the cost and time efficiency of AI algorithms. Van Gasteren enthusiastically explains that they are currently working hard to help clients apply it across a variety of sectors and applications.