Optimal Multi-View Surface Normal Estimation Using Affine Correspondences.

Area of Science:

• Computer Vision
• Photogrammetry
• Computer Graphics

Background:

• Accurate surface normal estimation is crucial for 3D reconstruction.
• Existing methods often rely on geometric information or are computationally expensive.
• Photometric stereo and multi-view stereo (MVS) methods have limitations in accuracy and speed.

Purpose of the Study:

• To propose an optimal least-squares method for estimating surface normals using exclusively photometric information.
• To develop a robust extension for outlier elimination.
• To demonstrate the method's effectiveness and applications in 3D reconstruction.

Main Methods:

• Exploiting photometric information via affine correspondences.
• Estimating surface normals independently as roots of a quartic polynomial.
• Employing maximum likelihood estimation and iteratively re-weighted least squares for robustness.

Main Results:

• The proposed method achieves superior accuracy and processing time compared to state-of-the-art techniques.
• Validated on both synthetic and real-world datasets.
• Demonstrated significant improvements in 3D reconstruction accuracy (30% reduction in 3D point error) when used in MVS.

Conclusions:

• The novel photometric approach provides an efficient and accurate solution for surface normal estimation.
• The robust extension effectively handles outliers, enhancing reliability.
• The method shows significant potential for improving multi-view stereo reconstruction and oriented point cloud applications.