Analyzing the effects of different 3D-model acquisition methods for synthetic AI training data generation and the domain gap

First published in
Lecture Notes in Mechanical Engineering
Number in series
Start Page
149
End Page
159
Citation
32nd International Conference on Flexible Automation and Intelligent Manufacturing, FAIM 2023
Contribution to Conference
32nd International Conference on Flexible Automation and Intelligent Manufacturing, FAIM 2023  
Publisher DOI
10.1007/978-3-031-38241-3_18
Scopus ID
2-s2.0-85171548382
Publisher
Springer Nature Switzerland
ISBN
978-3-031-38240-6
978-3-031-38241-3
978-3-031-38242-0
Synthetic data generation to enable industrial visual AI applications is a promising alternative to manual data acquisition. Such methods rely on the availability of 3D models to generate virtual worlds and derive or render 2D image data. Since CAD models can be created on different detail levels but require extensive effort, questions arise about which affect different detail levels of 3D models have on synthetic data generation. The effect of different CAD model details is investigated and compared to different 3D scanning and photogrammetry approaches. Different synthetic datasets are created for each 3D model acquisition type and a test-benchmark set is used to evaluate the performance of each dataset. Based on the results, the suitability of each acquisition method is derived and the effects of bridging the domain gap from the synthetic training domain to the real-world application domain are discussed. The findings indicate that 3D scans are as feasible for synthetic data generation as feature-rich high-level CAD data. Feature poor CAD data as they might originate from manufacturing data performs significantly worse.
Subjects
3D Scanning
Data Acquisition
Domain Gap
Photogrammetry
Synthetic Data Generation
DDC Class
004: Computer Sciences
620: Engineering