Collocating Clothes With Generative Adversarial Networks Cosupervised by Categories and Attributes: A...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Clothing choices impact personal perception and style expression.
• Conditional Generative Adversarial Networks (cGANs) have advanced image-to-image translation.
• Existing cGAN frameworks typically use a single generator and discriminator.

Purpose of the Study:

• To propose a novel cGAN framework utilizing multidiscriminators for improved clothing matching.
• To investigate the efficacy of incorporating diverse conditional information into discriminators.
• To enhance the quality and relevance of generated clothing images.

Main Methods:

• Developed a new framework with multiple discriminators for cGANs.
• Introduced Attribute-GAN (two discriminators) and Category-Attribute GAN (CA-GAN, three discriminators).
• Created a large-scale dataset of 19,081 clothing image pairs with 90 labeled attributes.

Main Results:

• The proposed multidiscriminator framework consistently improved generated clothing image quality.
• Attribute-GAN and CA-GAN demonstrated superior performance compared to state-of-the-art methods.
• Supervision from additional attribute or category discriminators enhanced GAN performance.

Conclusions:

• Multidiscriminator cGAN frameworks offer a promising direction for generating high-quality clothing matches.
• Incorporating diverse conditional information significantly boosts the performance of generative models for fashion applications.
• The developed models and dataset provide valuable resources for research in fashion AI and computer vision.