Fast Photometric Stereo by Time- and Spectral-Multiplexing With Crosstalk Handling.

Area of Science:

• Computer Vision
• Computational Imaging
• Surface Reconstruction

Background:

• Photometric stereo is crucial for detailed surface normal recovery but faces accuracy-efficiency trade-offs.
• Conventional methods are accurate but slow; multispectral approaches are fast but suffer from crosstalk, harming accuracy.

Purpose of the Study:

• To resolve crosstalk in multispectral photometric stereo for fast and accurate surface normal estimation at low cost.
• To develop a framework for Fast Photometric Stereo (FPS) that addresses crosstalk limitations.

Main Methods:

• Analysis of crosstalk formulation and impact, identifying external and internal factors.
• Proposed FPS framework with a combined time- and spectral-multiplexing data capture scheme.
• Introduction of FPS-Net, a lightweight network for crosstalk removal based on derived constraints.

Main Results:

• Demonstrated that external factors are primary contributors to crosstalk, followed by internal factors.
• Developed a real-world dataset affected by crosstalk for evaluation.
• Experimental results show superior accuracy and efficiency of the proposed FPS method over existing techniques.

Conclusions:

• The proposed FPS framework effectively resolves crosstalk issues in photometric stereo.
• Achieved a balance between accuracy and efficiency in surface normal estimation.
• The FPS-Net model provides a computationally efficient solution for crosstalk removal.