Dense photometric stereo: a Markov random field approach.

Area of Science:

• Computer Vision
• Computer Graphics
• Computational Imaging

Background:

• Photometric stereo is crucial for 3D surface reconstruction but is sensitive to complex scene properties.
• Existing methods struggle with challenges like complex geometry, shadows, highlights, and variable lighting conditions.

Purpose of the Study:

• To develop a robust normal reconstruction method using dense photometric stereo.
• To improve accuracy and efficiency in surface normal estimation and enable real-time relighting.

Main Methods:

• Formulating dense photometric stereo as a Markov Random Field (MRF) problem.
• Applying graph cuts and tensor belief propagation for Markov Random Field (MRF) inference.
• Developing a surface reconstruction algorithm from the recovered normal map.
• Implementing real-time relighting using graphics processing units (GPUs).

Main Results:

• Achieved robust and faithful normal reconstruction for complex scenes, preserving discontinuities.
• Tensor belief propagation demonstrated faster convergence and reduced computational requirements compared to graph cuts.
• Successfully reconstructed surfaces and enabled real-time novel view synthesis under arbitrary lighting conditions.

Conclusions:

• The proposed MRF-based photometric stereo method offers a significant advancement in robust normal reconstruction.
• The developed relighting technique provides a powerful tool for novel image synthesis with high efficiency.
• The method effectively handles challenging imaging conditions and complex scene geometries.