PoT-GAN: Pose Transform GAN for Person Image Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Pose-based person image synthesis generates new images of a person in a target pose from a source image.
• Significant pose differences cause appearance discrepancies, posing a challenge for existing methods.

Purpose of the Study:

• To present the Pose Transform Generative Adversarial Network (PoT-GAN) for improved person image synthesis.
• To enable reliable appearance transfer to arbitrary target poses without hard-coded spatial information.

Main Methods:

• The proposed PoT-GAN generator explicitly learns pose transformations.
• It manipulates multi-scale feature maps to incorporate learned pose transform information.
• A Generative Adversarial Network (GAN) architecture is utilized.

Main Results:

• PoT-GAN successfully transfers appearance from source to target poses.
• The method demonstrates state-of-the-art performance on three public datasets.
• Qualitative and quantitative evaluations confirm its effectiveness.

Conclusions:

• PoT-GAN offers a robust solution for pose-based person image synthesis.
• The explicit learning of pose transformations addresses key challenges in appearance transfer.
• The model achieves superior performance compared to existing approaches.