Non-Lambertian Photometric Stereo Network based on Inverse Reflectance Model with Collocated Light

Area of Science:

• Computer Vision
• Computer Graphics
• Photometry

Background:

• Traditional photometric stereo requires numerous images for accurate surface normal estimation.
• Non-Lambertian reflectance poses challenges for conventional methods.

Purpose of the Study:

• To develop a non-Lambertian photometric stereo network for accurate surface normal recovery with sparse lighting.
• To address limitations of current methods requiring a large number of images.

Main Methods:

• Derived an inverse reflectance model using monotonicity and univariate properties for collocated light.
• Employed supervised deep learning for enhanced shadow rejection and model flexibility.
• Utilized max-pooling for shadow handling and neighborhood image patches for reflectance variation.

Main Results:

• The proposed method achieves state-of-the-art accuracy in surface normal estimation.
• Demonstrated effectiveness on both synthetic and real-world image datasets.
• Successfully decouples surface normal from reflectance, enabling sparse light recovery.

Conclusions:

• The novel network provides a robust solution for non-Lambertian photometric stereo under sparse lighting conditions.
• Deep learning integration significantly enhances shadow rejection and adaptability to diverse reflectances.
• The method offers a more efficient and accurate approach to surface normal estimation.