Revisiting Supervised Learning-Based Photometric Stereo Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Photometric Stereo

Background:

• Deep learning advances photometric stereo, yet its mechanisms for handling unknown reflectance and global illumination are unclear.
• Supervised learning methods in photometric stereo face challenges in feature representation and network design.

Purpose of the Study:

• To elucidate how supervised deep learning methods address challenges in photometric stereo.
• To propose an effective network architecture for enhanced normal estimation in photometric stereo.

Main Methods:

• Analysis of deep features, encoding strategies, and network architectures in existing photometric stereo methods.
• Development of ESSENCE-Net featuring an easy-first-encoding strategy for shading features.
• Integration of shading supervision and spatial context-aware attention for accurate normal decoding.

Main Results:

• ESSENCE-Net demonstrates superior performance compared to state-of-the-art methods.
• The proposed method achieves high accuracy on benchmark datasets with both dense and sparse inputs.
• Validation of the effectiveness of the easy-first-encoding strategy and spatial context-aware attention.

Conclusions:

• ESSENCE-Net provides an effective solution for photometric stereo by optimizing feature encoding and decoding.
• The insights gained from analyzing existing methods inform the design of improved photometric stereo networks.
• The proposed approach advances the field of photometric stereo by enhancing normal estimation accuracy.