SYNTHETIC ASSET CREATION FOR OBJECT RECOGNITION
SORDI
Year//
2022
Direct Team//
Project Partners//
SORDI (Synthetic Object Recognition Dataset for Industries) accelerates artificial intelligence in production pipelines. The Synthetic dataset contains over 800,000 photorealistic images in 80 categories for production and contributes to the accessibility of tangible training data for AI production. The visual data uses advanced rendering technology and allows the integrated digital assets to enable image processing tasks such as classification, object detection and segmentation for areas of manufacturing, quality control and extended areas of the production pipeline. SORDI increases the efficiency of training workflows, further strengthening no-code AI solutions and will help establish newer robust AI models.
PHOTOREALISM AT ITS CORE
DIGITAL ASSETS
During my 2020 internship at BMW and Idealworks, our team had the opportunity to convert an assortment of factory assets into digital twins. At the core of this process was a proven Physically based rendering (PBR) material workflow, commonly employed in game engines such as Unreal and Unity. We employed this method to mimic factory conditions and manually texture details such as dust, scratches and general wear - a feature that was eventually built natively into the materials as a procedural function.
DYNAMIC MATERIALS FOR SIMULATIONS
PARAMETRIC
The dynamic element of the materials is based on depth, vertex density and procedural noise - elements that were manually accessible to our team during the production phase. When rebuilt within Omniverse however additional features such as interaction-based wear-and-tear were introduced - enabling the assets to dynamically change based on environmental factors and use-case scenarios. In light of the publication of the dataset, we created a short animation reel to highlight the parametric nature of the assets:
LAYOUTS AND BUILT ENVIRONMENTS
THE iFACTORY
After a sizable portion of the required assets was completed, we devoted some resources to positioning assets within the bounds of a mapped layout. We experimented with the beta version of Nvidia Nucleus and used Unreal and Blender to build out different factory levels during workflow case studies. Our level designs and layouts would eventually culminate as the iFactory, a functional digital twin pulling data from manufacturing workflows and visualising the plant’s operations within Nvidia’s Omniverse.
RESEARCH OUTCOMES
AWARNESS
SORDI assets were used to demonstrate the framework within an industrial setup to build a case for situational awareness as a fundamental element for organizing and managing robots within future smart factories. The details of this paper can be found below. This research was undertaken by idealworks in collaboration with Boulos Asmar, and aimed to convert low-level data streams into high-level semantic representations to build a case for autonomous vehicle management methods.
THE LARGEST SYNTHETIC DATASET
DIGITAL FRONTIER
SORDI streamlines and accelerates the training of artificial intelligence and will allow companies to reach higher levels of autonomy within their production lines. Together with Microsoft, NVIDIA, and idealworks, SORDI has been made available as open-source, to build the world’s largest reference dataset for artificial intelligence in the field of manufacturing. The publication of this innovative dataset now represents the next step in BMW Group’s expansion of democratised artificial intelligence [https://github.com/bmw-innovationlab] and will continue to expand and include new models as the industry adapts to this new digital frontier
