Hierarchical Generation of Human Pose With Part-Based Layer Representation.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Human pose transfer is an emerging research area with unsatisfactory state-of-the-art results.
• Existing end-to-end methods lack semantic understanding, leading to limitations.

Purpose of the Study:

• To propose a novel human pose transfer method that incorporates semantic part-based representations.
• To improve the quality and handling of self-occlusions in human pose transfer.

Main Methods:

• Segmenting the human body into multiple semantic parts.
• Utilizing part-based layer generators for high-quality local region results.
• Designing a three-stage hierarchical framework for coarse-to-fine fusion of local representations.
• Exploiting spatial guidance from 3D human models to handle self-occlusions.

Main Results:

• Achieved superior quantitative and qualitative performance compared to previous approaches.
• Demonstrated improved handling of self-occlusions, reducing artifacts.
• Showcased enhanced fine detail preservation and synthesis.
• Generated higher-resolution human pose transfer results.

Conclusions:

• The proposed semantic part-based and spatial-aware approach significantly advances human pose transfer.
• This method effectively addresses limitations of previous techniques, particularly in complex scenarios like self-occlusions.
• The hierarchical framework ensures both global consistency and local detail accuracy.