Surface-from-gradients without discrete integrability enforcement: A Gaussian kernel approach.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Geometry

Background:

• Existing surface reconstruction algorithms often enforce discrete integrability, leading to limitations like handling only dense fields, smoothing sharp features, or causing surface distortion.
• Traditional methods frequently employ Fourier or wavelet bases, which are prone to discretization and finite approximation errors, resulting in inaccurate surface representations.

Purpose of the Study:

• To develop a new 3D surface reconstruction method that avoids discrete integrability constraints and mitigates common issues in existing techniques.
• To reconstruct continuous 3D surfaces from potentially sparse or dense gradient and/or height fields with enhanced accuracy and feature preservation.

Main Methods:

• A novel kernel-based approach is proposed, transferring the continuous surface reconstruction problem into a high-dimensional space.
• Utilizes kernel basis functions, specifically the Gaussian kernel, to derive a closed-form solution for surface reconstruction.
• The method does not enforce discrete integrability, thereby avoiding discretization and finite approximation errors.

Main Results:

• The proposed kernel-based method achieves accurate 3D surface reconstruction, outperforming traditional techniques.
• Demonstrates superior preservation of salient and sharp features compared to classical and recent methods.
• Successfully reconstructs surfaces from both dense and sparse gradient or height fields.

Conclusions:

• The kernel-based surface reconstruction method offers a significant advancement over existing approaches by avoiding discrete integrability.
• The technique provides accurate and distortion-free surface reconstruction, preserving critical details.
• The availability of source code facilitates further research and application of this method.