UVIRT-Unsupervised Virtual Try-on Using Disentangled Clothing and Person Features

Area of Science:

• Computer Vision
• Artificial Intelligence
• E-commerce Technology

Background:

• Virtual try-on (VTON) is crucial for the e-commerce industry, enabling realistic clothing superposition on individuals.
• Existing VTON methods predominantly rely on supervised learning, requiring expensive annotated data (e.g., segmentation masks, paired images).
• Weakly-supervised VTON methods also incur high annotation costs or utilize pre-trained networks, which can introduce latency issues in practical applications.

Purpose of the Study:

• To develop a novel unsupervised virtual try-on method that eliminates the need for annotated data.
• To address the high costs and latency associated with existing supervised and weakly-supervised VTON approaches.
• To enable efficient and cost-effective virtual try-on solutions for e-commerce.

Main Methods:

• Propose Unsupervised VIRtual Try-on using disentangled representation (UVIRT).
• Extracts distinct features for clothing and person from respective images.
• Exchanges these disentangled features to achieve virtual try-on without supervision.

Main Results:

• UVIRT achieves comparable results to supervised methods on paired image datasets (MPV).
• UVIRT outperforms supervised methods on unpaired image datasets (C2C marketplace).
• Demonstrates the effectiveness of unsupervised learning for virtual try-on.

Conclusions:

• UVIRT offers a cost-effective and efficient alternative to supervised virtual try-on methods.
• The unsupervised approach removes the dependency on annotated data, pre-trained networks, and category labels.
• UVIRT shows significant potential for practical applications in the e-commerce sector, especially with unpaired data.