Pose-Driven Realistic 2-D Motion Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computer Graphics

Background:

• Realistic 2-D motion synthesis is crucial for animation and virtual environments.
• Existing methods struggle with spatial deformations inherent in motion synthesis.
• Generative Adversarial Nets (GANs) show promise but are limited to region-aligned tasks.

Purpose of the Study:

• To develop a novel approach for realistic 2-D motion synthesis.
• To overcome the limitations of GANs in handling spatial deformations in motion synthesis.
• To generate high-fidelity motion images conditioned on human poses.

Main Methods:

• Proposing a two-step, multistream network architecture for motion synthesis.
• Utilizing pose estimation technology and GANs for image generation.
• Generating body segment images in step-I and textures in aligned regions in step-II.
• Incorporating a real face as input to enhance facial details.

Main Results:

• Successfully synthesized realistic 2-D motion with sharp details.
• Demonstrated the effectiveness of the proposed model on four training sets.
• The novel architecture effectively addresses spatial deformation challenges in motion synthesis.

Conclusions:

• The proposed method transforms 2-D motion synthesis into a pose-conditioned image generation task.
• The two-step, multistream network architecture effectively handles spatial deformations.
• The model generates realistic and detailed motion images, outperforming previous approaches.