Surface Reconstruction from Structured Light Images Using Differentiable Rendering.

Area of Science:

• Computer Vision
• Geometric Modeling
• 3D Reconstruction

Background:

• Traditional 3D scanning methods generate point clouds, requiring a separate surface reconstruction step.
• Errors in point cloud computation can negatively impact the final surface reconstruction quality.
• Existing methods are often sensitive to noise and may struggle with sharp features.

Purpose of the Study:

• To develop a direct, one-step surface reconstruction method from structured light images.
• To improve the accuracy and robustness of 3D surface reconstruction.
• To overcome limitations associated with intermediate point cloud representations.

Main Methods:

• A novel approach directly computes a triangle mesh surface from structured light images.
• Minimizes least-squares error between input images and renderings of the triangle mesh.
• Utilizes differentiable rendering for efficient optimization and gradient computation.

Main Results:

• The proposed method achieves accurate 3D surface reconstructions, outperforming point cloud-based approaches.
• Demonstrates superior reconstruction of sharp edges and robustness against image noise.
• Shows potential for improving existing reconstruction algorithms through initialization.

Conclusions:

• Direct surface computation from structured light images offers significant advantages over traditional methods.
• The differentiable rendering approach enables efficient and accurate 3D surface reconstruction.
• This method provides a robust and versatile tool for 3D scanning and modeling.