CADaVISION: A simulation framework for machine vision prototyping

Abstract

This paper describes a system that e\\'0eciently creates synthetic ground-truth data to support the engineering of machine vision systems while keeping requirements to the user lower than using standard 3D-modeling software. Utilizing the power and exibility of programmable GPUs, image rendering is carried out with a high degree of realism while still achieving high frame rates for real-time camera view simulation. For images requiring a very high level of detail in reection and global illumination calculation, the framework includes the possibility to use ray tracing based rendering. Using the synthetic images, processing algorithms are selected and parameterized. Results of the processing algorithms are then compared between synthetic images based on CAD-models and real-world scenes of single workpieces and assemblies. To aquire reection properties of materials and to capture images of workpieces and assemblies under varying illumination, a robot- based gonioreectometer was set up. Our system exchanges data between components using COLLADA.