Industrial object detection, pose estimation and surface inspection are common industrial tasks. In Flanders Make we develop Artificial Intelligence (AI) techniques that can be trained fully using Synthetic data by rendering photorealistic images, with their embedded annotations, starting from Computer-Aided Design (CAD) models of these objects.
Three toolboxes have been developed, respectively, CAD2DETECT, CAD2POSE and CAD2DEFECT. These toolboxes are all end-to-end solutions for use in industrial environments to train and deploy, for instance, neural networks. The synthetic data generation is done using another Flanders toolbox called CAD2RENDER, which is a fully customisable dataset generation toolbox, which uses only a CAD file, information on the surface texture and domain knowledge of an object as input for the generation of huge, fully annotated datasets with depth maps. Altogether, they solve the need for large, annotated image datasets which are difficult to produce as it is hard, repetitive work to generate the annotations, prone to errors, if done using real images. The synthetic images can be generated in a fraction of the time compared to the traditional methods. The full flow for training and deploying these models is illustrated in the graph below.
We can describe these 3 toolboxes in terms of the industrial applications in which these can be:
CAD2DETECT is a tool allowing for fast and accurate object detection, identification and counting. A typical usage of this toolbox is sorting of various objects on a high-speed conveyor belt, based on the type of object. The tool uses a simple RGB camera for inference, it can potentially distinguish between hundreds of objects and deal with object occlusions.
CAD2POSE is a solution for 6DOF pose estimation tasks starting from a standard RGB camera, such as needed in a robotics bin-picking application. Highly robust against occlusion and harsh lighting conditions, it can accurately predict the pose of all kinds of objects, even if they are very flat, symmetrical, or small. In combination with the CAD2DETECT toolbox, it is incredibly useful for managing mixed object streams in either stationary or dynamic sorting and robotic (bin) picking applications.
The CAD2DEFECT solution is a highly accurate and fast surface defect detection tool for quality assurance. The initial state of the network is trained on a large synthetic dataset where virtual defects are present. The model is then further trained in a different way using a small dataset of real images containing no objects with defects, which are 99.9% of objects in the industry. This makes dataset generation very easy and efficient. This solution provides robust defect detection and furthermore, reduces the deployment time of such detection in an industrial setting significantly compared to other approaches.
For more information, please contact: