Fully automated structured light scanning for high-fidelity 3D reconstruction via graph optimization

Area of Science:

• Computer Vision
• Geometric Modeling
• 3D Reconstruction

Background:

• Current 3D scanning methods face challenges with high manual costs and error accumulation in large-scale projects.
• Achieving industrial-grade, seamless, and high-fidelity 3D models with minimal manual intervention remains a significant challenge.

Purpose of the Study:

• To develop an innovative method for industrial-grade seamless and high-fidelity 3D reconstruction with minimal manual intervention.
• To address error accumulation in multi-frame registration and improve surface model quality.

Main Methods:

• Transformed multi-frame registration into a graph optimization problem using point cloud cross-multipath registration for global consistency.
• Employed dynamic nonlinear weights based on geometric and color differences, optimized using iteratively reweighted least squares (IRLS) for bundle adjustment (BA).
• Introduced a local-to-global transitioning strategy for multi-frame fusion to ensure watertight, seamless surface models and correct normal vector consistency.

Main Results:

• Significantly enhanced registration accuracy and robustness compared to traditional frame-by-frame methods.
• Achieved near real-time efficiency while maintaining high fidelity in 3D reconstruction.
• Successfully generated watertight and seamless surface models, effectively correcting mesh discontinuities.

Conclusions:

• The proposed graph optimization and multi-frame fusion strategy offers a robust and efficient solution for industrial-grade 3D reconstruction.
• The method demonstrates superior performance in accuracy, robustness, and seamless surface generation.
• The open-sourced data and implementation encourage further community development in 3D reconstruction technologies.