Expertise
synthetic training data for machine vision.
Imagine a pipe inspection robot. How do you train its vision algorithm when neither the robot nor the piping network has been built yet? In other use cases, some real-life data may be available, but the dataset is imbalanced, critical edge cases are missing or annotation is lacking. A large set of diverse, well-annotated and relevant data is key to developing a successful machine learning algorithm. Synthetic training data offers a solution where this data is not available or incomplete. Demcon develops project-specific simulations which generate near-infinite permutations of specific 3D environments. These environments are sampled by a range of virtual sensors to generate rich, well-balanced datasets.
highlights
- Digital Twin of project-specific 3D environment generates rich datasets and annotation
- Procedural methods create near-infinite permutations of objects and environments
- Pixel perfect annotation
- Simulation of a range of sensors, e.g. RGB(D), infrared, Lidar, et cetera
- Rare edge cases can be generated at will to balance the dataset
- Enables Machine Learning where data or accurate annotation is not obtainable
- Develop in silico: digital twin of environment, device and sensors
a known ground truth provides perfect annotation.
A simulation provides a known ground truth, enabling rich, pixel-perfect annotation. This includes annotation which is typically very hard or impossible to do, such as segmentation, depth, surface normals, optical flow, consistent quality indicators on a per-pixel or per-object basis and much more.
variation beyond the basics.
Demcon develops procedural models and methods to construct our project-specific environments. This means our models include methods that are developed to enable variation beyond basic parameters such as color, size and orientation.
Organic subjects such as fruits and vegetables are inherently difficult to capture due to their large natural variation. Our approach is not to model a great number of different models but to develop one procedural model. The parametric interface to these procedures enables the generation of near-infinite variations by varying e.g. shape parameters, age and defects. By linking distributions of parameters, we can develop high-level interfaces to generate random instances of fruit of arbitrary classifications or quality levels. This allows for annotation of highly consistent quality indicators on both a pixel-level and an object-level throughout the system.
all expertises.
