Reproduce “Global and Local Texture Randomization for Synthetic-to-Real Semantic Segmentation” on OpenResearch

About this paper

Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, Jun Liu

Researchers developed a framework employing global and local texture randomization with "unreal" painting textures to enhance the generalization of semantic segmentation models trained on synthetic data to unseen real-world environments. The approach consistently improved mean Intersection over Union (mIoU) across various benchmarks, outperforming existing domain generalization methods and achieving results comparable to several domain adaptation techniques.

Key insights

• Randomizing textures using "unreal" painting styles, rather than real-world styles, effectively desensitizes CNNs to texture bias, promoting the learning of domain-invariant shape and semantic features.
• Domain shifts often manifest locally, and specific randomization of local textures (LTR) offers benefits beyond global style transfer for improved generalization.
• Enforcing consistency between global and local texture randomization (CGL) mechanisms is crucial to ensure harmonious learning and prevent conflicting gradients.

Summary by alphaXiv.