Using machine learning for computer vision applications is extremely time consuming since many pictures need to be taken and labelled manually.
Our approach eliminates this expensive process by using synthetic renderings and artificially generated pictures for training. The main challenge is to overcome the visual difference between artificial renderings and camera pictures. This difference is called the domain gap and it rendered preceding attempts to train neural camera networks with synthetic data useless.
By modifying the artificial renderings in such a way that the difference to real camera pictures is minimized, we were able to close this domain gap therefore allowing us to train computer vision neural networks with synthetic data.
This enables us to detect and classify objects in pictures by training a neural network with renderings of the respective 3D models. Rendering these objects from various perspectives yields our training data set. Being able to use artificially generated pictures to train a neural network, that works with real world pictures, basically allows us to train any image-based supervised model.
The proposed technology is a software product that is able to modify synthetic renderings in such a way that they resemble camera pictures and can be used for training.
Therefore, when given a 3d model, our technology is able to generate thousands of renderings within seconds showing the object from different perspectives. Subsequently, the object can be detected, classified and examined in camera pictures. The same process would take hours for every object when conducted manually.
Closing the domain gap between camera images and synthetic renderings is our accomplishment.
Our technology is useful for every neural network that analyzes camera pictures, since we are able to create training data sets artificially. This vastly increases efficiency.
Computer Vision tasks are performed in various industries. In the manufacturing industry objects are often captured by cameras in order to guarantee specific quality standards. The captured object is detected, classified and automatically examined for flaws or defects. Also the classification process allows sorting tasks to be executed.
Our approach works regardless of the texture of the actual object since our training method ensures that the neural network is able to detect the subject's geometric shape rather than its texture. It is therefore also possible to create a coherent texture from multiple camera images.
Basically any algorithm that depends on a training set of camera pictures will benefit from our technology since the training set can be extended vastly by using synthetic data.
By using our technology, customers are able to use machine learning-based computer vision applications. In contrast to competitors, we are able to offer specifically tailored systems for a lower price since there is no need for time-consuming and expensive creation of training datasets.
We are lowering the bar for every company to access the world of artificial intelligence and machine learning procedures.
Even proof-of-concepts and prototypes for computer vision neural networks can be implemented and tested fast. This technology offers huge potential for automation of quality control, sorting and texturization tasks.