Synthetic data generation procedures for domain-specific environments in manufacturing

Synthetic data is being used lately for training deep neural networks in computer vision applications such as object detection, object segmentation and 6D object pose estimation. Domain randomization hereby plays an important role in reducing the simulation to reality gap. However, this generalization might not always be effective in specialized domains like manufacturing that involve complex assemblies. Individual parts are integrated in much larger assemblies making them indistinguishable from their counterparts. Moreover, individual parts are often partially occluded in the scene. These situations give rise to wrong detections. Target domain knowledge is vital in these cases and if conceived effectively while generating synthetic data, can show a considerable improvement in bridging the simulation to reality gap. This paper validates synthetic data generation procedures through practical experimentation ensuring that experiments are both comprehensive and reproducible. After combining domain randomization and domain adaptation procedures for parts and assemblies used in manufacturing the model performance improves by up to 15% than the state-of-the-art domain randomization techniques. Reducing the simulation to reality gap in this way can unlock the true potential of robot-assisted production using artificial intelligence.

Subjects

Domain knowledge | Object detection | Sim2Real gap | Synthetic data

DDC Class

006: Special computer methods

670: Manufacturing

681: Precision Instruments and Other Devices

Publication version

publishedVersion

Lizenz

https://creativecommons.org/licenses/by-nc-nd/4.0/

Name

1-s2.0-S1877050925002376-main.pdf

Type

Main Article

Size

1.4 MB

Format

Adobe PDF

Options

Synthetic data generation procedures for domain-specific environments in manufacturing