Deep Photometric Stereo for Non-Lambertian Surfaces.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Photometric stereo traditionally relies on simplified surface reflectance models.
• Handling non-Lambertian surfaces and uncalibrated lighting presents significant challenges.

Purpose of the Study:

• To develop a deep learning-based photometric stereo method for both calibrated and uncalibrated scenarios.
• To address the limitations of traditional methods in handling complex surface reflectance and unknown lighting conditions.

Main Methods:

• Introduced PS-FCN, a fully convolutional network for calibrated photometric stereo, learning direct mapping from observations to surface normals.
• Developed LCNet to estimate unknown light directions in uncalibrated scenarios, enabling subsequent normal estimation via PS-FCN.
• The approach handles general, unknown isotropic reflectance and is order-agnostic to input images.

Main Results:

• PS-FCN directly learns reflectance-to-normal mapping, surpassing traditional simplified models.
• LCNet successfully estimates light directions, allowing PS-FCN to work in uncalibrated settings.
• The combined method demonstrates superior performance over state-of-the-art techniques on synthetic and real-world data.

Conclusions:

• Deep learning offers a robust solution for photometric stereo on non-Lambertian surfaces.
• The proposed PS-FCN and LCNet networks effectively handle both calibrated and uncalibrated conditions.
• This work advances surface reconstruction by enabling accurate normal estimation with complex materials and unknown lighting.