PEAR: Pixel-aligned Expressive humAn mesh Recovery

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• PEAR achieves inference speeds over 100 FPS for SMPLX and scaled-FLAME parameter estimation.result 90%
• PEAR significantly improves pose estimation accuracy compared to previous SMPLX-based methods.result 85%
• PEAR uses a clean and unified ViT-based model to recover coarse 3D human geometry for real-time inference.method 90%
• PEAR introduces pixel-level supervision to optimize geometry and improve reconstruction accuracy of fine-grained details.technical 90%
• Existing SMPLX-based methods often suffer from slow inference, coarse body poses, and misalignments in fine-grained regions.limitation 85%
• PEAR proposes a modular data annotation strategy to enrich training data and enhance model robustness.method 85%
• PEAR explicitly addresses inaccurate localization of fine-grained human pose details and insufficient facial expression capture.technical 90%
• Potential limitations include reliance on specific ViT architectures and challenges in handling diverse ethnic and age datasets.limitation 80%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card