Joel Sol

Topic

A Sim-to-Real Deformation Classification Pipeline using Data Augmentation and Domain Adaptation

Department of Electrical and Computer Engineering

Date & location

• Wednesday, May 8, 2024

• 10:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Electrical and Computer Engineering, University of Victoria (Supervisor)

• Dr. Stephen Neville, Department of Electrical and Computer Engineering, UVic (Member)

External Examiner

• Dr. Caterina Valeo, Department of Mechanical Engineering, UVic 

Chair of Oral Examination

• Dr. David Leitch, Department of Chemistry, UVic

   

Abstract

Geometrical quality assurance is critical for improving manufacturing time and cost. This is more inhibiting when the visual or haptic assessment of human operators is necessary. Modern machine learning (ML) methods can solve this problem but require large datasets including diverse deformations. However, preparing those deformations using physical objects can be difficult and costly. This thesis proposes to use Blender, an open-source simulation tool, to imitate object deformities and automate the preparation of synthetic datasets. The utility of these datasets can be further improved using data augmentation techniques such as background randomization or domain adaptation networks. The background randomization approach provides a way to generalize the image distribution to a variety of environments, whereas the domain adapted approach provides a better targeted distribution. This thesis showcases that discrepancies between real and simulated environments can be mitigated to create models effective at sim-to-real deformation detection.