Fully Deformable Network for Multiview Face Image Synthesis.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Photorealistic multiview face synthesis from a single image is challenging.
• Existing methods struggle with large pose variations and facial detail loss.
• Geometric constraints in face deformation are often overlooked.

Purpose of the Study:

• To propose a novel fully deformable network for accurate face synthesis.
• To address limitations in handling large pose variations and preserving facial details.
• To explicitly model geometric constraints for improved face generation.

Main Methods:

• Developed a fully deformable network with a deformable landmark learning network (DLLN) and a gated deformable face synthesis network (GDFSN).
• DLLN guides pose rotation by converting reference landmarks to target landmarks.
• GDFSN uses a dual-stream structure to estimate and apply deformation offsets, incorporating a soft-gating mechanism and pose-aligned skip connections (PASC).

Main Results:

• The proposed method achieves state-of-the-art performance, particularly excelling in synthesizing faces with large pose changes.
• Experiments on six benchmarks demonstrate superior results compared to existing approaches.
• Enhanced facial details and identity preservation were observed.

Conclusions:

• The fully deformable network effectively models individual-aware pose changes and geometric transformations for face synthesis.
• The approach overcomes limitations of previous methods in handling pose variations and detail loss.
• The method offers a significant advancement in single-image based photorealistic face synthesis.