PoNA: Pose-guided Non-local Attention for Human Pose Transfer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Synthesis

Background:

• Human pose transfer is crucial for applications but often yields blurry results due to limitations in previous methods.
• Existing techniques often neglect pose features or rely solely on local attention, failing to capture global image context effectively.

Purpose of the Study:

• To develop an advanced human pose transfer method that overcomes the limitations of existing approaches.
• To generate high-fidelity, realistic human images with accurate pose representation and rich details.

Main Methods:

• A novel generative adversarial network (GAN) architecture featuring simplified cascaded blocks.
• Introduction of a pose-guided non-local attention (PoNA) mechanism for enhanced feature selection and long-range dependency modeling.
• Implementation of pre-posed and post-posed feature update strategies to optimize pose and image information integration.

Main Results:

• The proposed method achieves superior quantitative and qualitative results on Market-1501 and DeepFashion datasets.
• Generated images exhibit enhanced sharpness, realism, and detail compared to state-of-the-art methods.
• The model demonstrates efficiency with fewer parameters and faster processing speeds.

Conclusions:

• The developed human pose transfer method is effective, efficient, and stable.
• The approach successfully addresses challenges in pose transfer, producing high-quality visual outputs.
• Generated images can potentially mitigate data scarcity issues in person re-identification tasks.