DuetGS: Two-Stage Controllable 3D Human Reconstruction From Dual Images.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Reconstructing detailed 3D human models from limited views is challenging.
• Existing methods struggle with spatial consistency and missing color information from few images.

Purpose of the Study:

• To propose DuetGS, a novel two-stage pipeline for realistic 3D human reconstruction.
• To address challenges in spatial consistency and color completion using minimal input views.

Main Methods:

• DuetGS divides reconstruction into geometry and color stages.
• A neural network recovers a full-body mesh for spatial positioning.
• Adapted Gaussian Splatting with unsupervised color propagation fills in missing details.

Main Results:

• DuetGS successfully reconstructs detailed 3D human models from front and back images.
• The method demonstrates superior reconstruction accuracy and visual quality compared to existing approaches.
• Gaussians mapped to the mesh allow for animation compatibility.

Conclusions:

• DuetGS offers an effective solution for 3D human reconstruction from limited views.
• The pipeline enhances spatial consistency and color detail.
• The approach shows significant improvements on benchmark datasets.