Geometry-Consistent Generative Adversarial Networks for One-Sided Unsupervised Domain Mapping.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain mapping aims to translate images between domains without paired data, which is an ill-posed problem.
• Existing methods like cycle consistency and distance preservation have limitations in constraining the solution space.
• These methods often overlook the inherent property of images where geometric transformations preserve semantic structure.

Purpose of the Study:

• To develop a novel approach for one-sided unsupervised domain mapping.
• To introduce a geometry-consistent generative adversarial network (Gc-GAN) that leverages geometric properties for improved image translation.
• To constrain the solution space effectively while retaining correct solutions in unsupervised domain mapping.

Main Methods:

• Developed a geometry-consistent generative adversarial network (Gc-GAN).
• Gc-GAN takes an original image and its geometrically transformed counterpart as input.
• Introduced a geometry-consistency constraint to couple generated images in the new domain.

Main Results:

• The geometry-consistency constraint effectively reduces the solution space for unsupervised domain mapping.
• Quantitative and qualitative comparisons demonstrate the effectiveness of Gc-GAN.
• Gc-GAN shows superior performance compared to baseline GAN and state-of-the-art methods like CycleGAN and DistanceGAN.

Conclusions:

• Gc-GAN offers an effective solution for one-sided unsupervised domain mapping.
• Preserving geometric structure is a valuable constraint for improving image translation tasks.
• The proposed method advances the field of unsupervised image-to-image translation.