Bayesian Cycle-Consistent Generative Adversarial Networks via Marginalizing Latent Sampling.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative Adversarial Networks (GANs), particularly cycle-consistent GANs, enable domain mapping with unpaired data but face training instability and mode collapse.
• Existing cycle-consistent models struggle with stable training and generating diverse outputs.

Purpose of the Study:

• To propose a novel Bayesian cyclic model and integrated framework to improve interdomain mapping stability and output diversity.
• To address the limitations of traditional CycleGAN, such as mode collapse and lack of diversified results.

Main Methods:

• Developed a Bayesian CycleGAN by exploring full posteriors via latent variable sampling and optimizing with Maximum A Posteriori (MAP) estimation.
• Integrated a novel cyclic framework to enhance the Bayesian approach.
• Enabled image diversification by manipulating latent variables during inference.

Main Results:

• Achieved a 15% improvement in per-pixel accuracy for semantic segmentation on Cityscapes within the original framework.
• Improved per-pixel accuracy by 20% on Cityscapes using the integrated framework.
• Demonstrated superior performance in style transfer tasks (Monet2Photo) with significantly more diversified results compared to the original CycleGAN.

Conclusions:

• The proposed Bayesian CycleGAN offers enhanced stability and robustness against adversarial imbalance.
• The method successfully diversifies generated images, overcoming a key limitation of previous models.
• Bayesian CycleGAN represents a significant advancement in unpaired image-to-image translation and style transfer.