Robust Multiview Photometric Stereo Using Planar Mesh Parameterization.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Photometric stereo (PS) reconstructs 3D shapes from images under varying lighting.
• Multiview stereo (MVS) methods provide initial 3D geometry but often lack fine details.
• Uncalibrated PS methods are challenging due to unknown camera and lighting parameters.

Purpose of the Study:

• To develop a robust uncalibrated multiview photometric stereo (MVS-PS) method for high-quality 3D shape reconstruction.
• To overcome limitations of conventional methods, such as merging view-dependent surface normal maps.
• To achieve accurate reconstruction of complex surfaces, including those with specular reflections.

Main Methods:

• A coarse 3D mesh from MVS is parameterized onto a 2D planar domain.
• Surface normal estimation is performed in the 2D space, integrating geometric and photometric cues.
• A refined 3D mesh is recovered via optimal displacement map computation in the 2D domain.

Main Results:

• The method successfully reconstructs accurate 3D meshes with fine geometric details.
• It handles challenging surfaces with specular reflections, multiple albedos, and complex topologies.
• Evaluations demonstrate robust performance in both controlled and uncontrolled environments.

Conclusions:

• The proposed uncalibrated MVS-PS method offers a robust and effective approach for high-quality 3D shape reconstruction.
• Parameterizing meshes to a 2D domain simplifies normal estimation and refinement.
• The technique avoids complex merging of view-dependent information, improving efficiency and accuracy.