Human as Points: Explicit Point-Based 3D Human Reconstruction From Single-View RGB Images.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Current single-view human reconstruction methods often rely on deep implicit functions, facing challenges in flexibility, generalizability, and robustness.
• Existing approaches struggle with controllable and unambiguous geometric representations, limiting their practical application.

Purpose of the Study:

• To introduce HaP, a novel explicit point-based framework for single-view human reconstruction.
• To overcome the limitations of implicit methods by proposing a fully explicit, geometry-centric approach.
• To provide a new benchmark dataset for real-scanned 3D human data.

Main Methods:

• HaP utilizes point clouds as an intermediate representation for 3D human geometry.
• The framework incorporates explicit point cloud estimation, SMPL manipulation, diffusion-based generation, and refinement through displacement learning and depth replacement.
• It operates entirely in the 3D geometric space, avoiding ambiguous implicit learning processes.

Main Results:

• HaP achieves quantitative performance improvements of 20-40% over state-of-the-art methods.
• The framework demonstrates superior qualitative results in human reconstruction.
• A new dataset of real-scanned 3D humans with intricate details is introduced.

Conclusions:

• The proposed HaP framework offers a more flexible, generalizable, and robust solution for single-view human reconstruction.
• Results suggest a potential paradigm shift towards fully-explicit, geometry-centric algorithms in this field.
• The new dataset will facilitate further research in detailed 3D human modeling.