Reproduce “Towards Sim-to-Real Industrial Parts Classification with Synthetic Dataset” on OpenResearch

About this paper

Xiaomeng Zhu, Talha Bilal, Pär Mårtensson, Lars Hanson, Mårten Björkman, Atsuto Maki

Researchers from Scania CV AB, University of Skövde, and KTH Royal Institute of Technology introduced SIP-17, a synthetic dataset designed for industrial parts classification, evaluating Sim-to-Real transfer by training models exclusively on synthetic data. The study found that domain randomization consistently enhanced classification accuracy, with the ConvNext model achieving 83.2% accuracy on isolated parts, while highlighting challenges in classifying highly similar or assembled industrial components.

Key insights

• Domain randomization, particularly the inclusion of random backgrounds and post-processing, significantly improves the Sim-to-Real transferability of models for industrial parts classification.
• State-of-the-art models like ConvNext and the self-supervised DINO exhibit strong capabilities for cross-domain classification in industrial settings.
• Classifying industrial parts with highly similar albedos, simple shapes, or subtle differences in assembled components presents considerable challenges for current Sim-to-Real approaches.

Summary by alphaXiv.