Progressive and Aligned Pose Attention Transfer for Person Image Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person pose transfer is a challenging task in computer vision.
• Existing methods often struggle with maintaining appearance and shape consistency.
• Data insufficiency is a common problem in tasks like person re-identification.

Purpose of the Study:

• To propose a new generative adversarial network (GAN) for accurate and photorealistic person pose transfer.
• To improve appearance and shape consistency in generated images.
• To demonstrate the utility of the proposed method for data augmentation in person re-identification.

Main Methods:

• A progressive generator architecture with a sequence of transfer blocks is designed.
• Attention mechanisms are employed within blocks to model pose relationships.
• Two novel blocks, Pose-Attentional Transfer Block (PATB) and Aligned Pose-Attentional Transfer Block (APATB), are introduced.

Main Results:

• The proposed model generates highly photorealistic person images.
• Superior appearance and shape consistency are achieved compared to existing methods.
• Quantitative and qualitative evaluations on Market-1501 and DeepFashion datasets validate the model's efficacy.

Conclusions:

• The developed GAN effectively performs person pose transfer with enhanced realism and consistency.
• The method serves as a valuable tool for data augmentation in person re-identification, mitigating data limitations.
• The research contributes to advancements in generative models for image manipulation and computer vision tasks.