Unsupervised Generation of Free-Form and Parameterized Avatars

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image-to-image translation is crucial for domain adaptation.
• Unsupervised learning methods are needed when paired data is unavailable.
• Generating realistic and controllable visual content remains a challenge.

Purpose of the Study:

• To develop unsupervised methods for mapping images between different domains.
• To extend image translation to generate a parameter vector and an image pair.
• To enable automatic avatar creation through image domain transfer.

Main Methods:

• Two novel unsupervised image-to-image translation frameworks were proposed.
• The methods were compared against unsupervised domain adaptation.
• Generative Adversarial Networks (GANs) were employed, guided by discrepancy-based generalization bounds.

Main Results:

• The proposed methods successfully perform unsupervised image domain transfer.
• The second method generates a tied pair of parameters and images.
• Experimental results validate the effectiveness for automatic avatar generation.

Conclusions:

• Unsupervised domain transfer can be effectively achieved for image mapping tasks.
• The developed GAN-based approach provides a viable solution for automatic avatar creation.
• This work advances unsupervised learning in computer vision with practical applications.