Reproduce “Mastering Sketching: Adversarial Augmentation for Structured Prediction” on OpenResearch

About this paper

Edgar Simo-Serra, Satoshi Iizuka, Hiroshi Ishikawa

This paper presents 'Adversarial Augmentation,' a framework that significantly advances sketch simplification by integrating supervised learning with unconditional adversarial training, allowing the use of both paired and unpaired data. The method achieves state-of-the-art performance in sketch cleanup, preferred by users 92.3% of the time, and enables new applications such as generating realistic pencil drawings from line art.

Key insights

• Unconditional discriminators in adversarial training allow for the effective incorporation of abundant unsupervised data, addressing limitations of purely supervised methods and improving generalization.
• Combining a pixel-wise supervised loss with an adversarial loss provides training stability and yields superior results for image-to-image translation tasks compared to using either loss in isolation.
• The framework's capacity to learn complex, non-deterministic mappings enables novel applications like generating realistic pencil drawings from line art, a task challenging for traditional pixel-wise methods.

Summary by alphaXiv.