Geometry-Constrained Non-Line-of-Sight Imaging.

Area of Science:

• Computer Vision
• Computational Imaging
• Geometry Processing

Background:

• Accurate surface and normal reconstruction is vital for non-line-of-sight (NLOS) imaging, enhancing scene understanding.
• Jointly estimating surface normals and albedo in NLOS imaging increases computational complexity due to tensor-valued functions.

Purpose of the Study:

• To develop a novel joint albedo-surface reconstruction method for NLOS imaging.
• To improve the accuracy of surface normal estimation for hidden objects using regularization techniques.
• To enhance the detail representation and geometric precision of reconstructed hidden objects.

Main Methods:

• Utilized the shape operator to regulate the normal field's variation rate.
• Applied regularization methods for the first time to reconstruct surface normals of hidden objects.
• Developed a joint albedo-surface reconstruction approach for NLOS imaging.

Main Results:

• The proposed method achieves high-precision reconstruction of hidden object geometry by improving normal field accuracy.
• Demonstrated robustness and effectiveness on both synthetic and experimental datasets.
• Achieved 30x speedup compared to existing methods on transient data, producing more accurate surfaces.

Conclusions:

• The surface normal-regularized reconstruction model offers superior accuracy and speed for NLOS imaging.
• This work represents a significant advancement in reconstructing hidden object geometry and appearance.
• The method enhances detail representation and provides a more comprehensive understanding of hidden scenes.