Three Dimensional Shape Reconstruction via Polarization Imaging and Deep Learning.

Area of Science:

• Computer Vision
• Optical Imaging
• Deep Learning

Background:

• Deep-learning 3D imaging estimates surface normals under passive lighting.
• Existing methods struggle with texture detail and normal accuracy due to information loss.

Purpose of the Study:

• To enhance texture detail restoration and surface normal estimation accuracy in 3D imaging.
• To mitigate information loss in fine-textured areas during object reconstruction.

Main Methods:

• Utilizing Stokes-vector-based parameters and separated reflection components for polarization input optimization.
• Developing networks to extract comprehensive polarization features and reduce background noise.

Main Results:

• The proposed method significantly improves surface normal estimation accuracy.
• Demonstrated reduction in mean angular error by 19% compared to UNet.
• Achieved 62% reduction in calculation time and 11% smaller model size.

Conclusions:

• The novel approach provides more accurate 3D reconstruction by enhancing polarization feature extraction.
• Optimized polarization representation leads to superior performance in surface normal estimation and texture recovery.