Unsupervised multi-domain multimodal image-to-image translation with explicit domain-constrained disentanglement

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image-to-image translation is a rapidly advancing field.
• Existing methods face challenges with aligned training data, output ambiguity, and multi-domain translation.
• A single network for simultaneous multi-domain translation is lacking.

Purpose of the Study:

• To propose a unified framework for image-to-image translation.
• To enable generation of diverse outputs using unpaired training data.
• To allow simultaneous multi-domain translation with a single model.

Main Methods:

• Developed a unified framework for image-to-image translation.
• Utilized unpaired training data for diverse output generation.
• Investigated explicit domain-level supervision for improved translation.

Main Results:

• The proposed method successfully generates diverse outputs from single inputs.
• Achieved simultaneous multi-domain translation using a single network.
• Demonstrated comparable or superior performance against state-of-the-art methods.

Conclusions:

• The unified framework effectively addresses key challenges in image-to-image translation.
• Explicit domain-level supervision enhances translation quality.
• The method shows broad applicability across various image translation tasks.