Not every sample is efficient: Analogical generative adversarial network for unpaired image-to-image translation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image-to-image (I2I) translation aims to map images between domains, with Generative Adversarial Networks (GANs) driving recent breakthroughs.
• Paired I2I methods require extensive input-output pairs, limiting practicality, while unpaired methods like CycleGAN use cycle-consistency on independent datasets.
• Existing unpaired I2I methods often use random sampling, neglecting sample similarity and potentially hindering optimal convergence.

Purpose of the Study:

• To develop a novel sampling strategy for unpaired image-to-image translation that enhances training efficiency and performance.
• To address the limitations of random sampling in dual learning-based I2I methods by incorporating content similarity.
• To introduce an analogical adversarial loss that leverages effective samples and mitigates the impact of less useful ones.

Main Methods:

• A novel metric-based sampling strategy inspired by analogical learning is proposed to select similar samples across domains for training.
• An analogical adversarial loss is introduced to guide the model towards learning from effective samples and reducing the influence of ineffective ones.
• The proposed framework is designed to be generic and easily extendable to existing I2I translation methods.

Main Results:

• Experimental results across various vision tasks demonstrate the superior performance of the proposed method compared to existing approaches.
• The new sampling strategy and loss function lead to more effective and efficient training for unpaired image translation.
• The generic nature of the framework allows for performance gains when applied to other I2I translation techniques.

Conclusions:

• The proposed metric-based sampling strategy and analogical adversarial loss offer a significant improvement for unpaired image-to-image translation.
• This approach enhances the effectiveness of GAN-based I2I translation by intelligently selecting training samples.
• The method provides a versatile framework applicable to various I2I tasks, promising broader advancements in the field.