Structure-guided adaptive total variation for parameter-free passive non-line-of-sight imaging.

Area of Science:

• Computer Vision
• Computational Imaging
• Signal Processing

Background:

• Non-line-of-sight (NLOS) imaging recovers scenes from indirect reflections.
• Existing NLOS methods struggle with adaptability and stability due to fixed or manual priors.
• Parameter sensitivity and color noise are persistent challenges in passive NLOS reconstruction.

Purpose of the Study:

• To develop a stable passive NLOS imaging framework using a conventional color camera.
• To overcome limitations of existing methods by introducing adaptive regularization.
• To enhance reconstruction quality and efficiency in diverse conditions.

Main Methods:

• Introduced a structure-guided adaptive total variation (SG-ATV) reconstruction framework.
• Developed a spatially varying regularization weight computation for automatic detail preservation and noise suppression.
• Utilized structural information from preliminary reconstruction to create a guidance map for robust optimization.
• Dynamically adjusted regularization weights iteratively, eliminating manual tuning.

Main Results:

• Achieved a 30x improvement in reconstruction efficiency.
• Significantly enhanced image quality, increasing PSNR from 17.52 dB to 58.92 dB.
• Reduced color deviations from 28.6 to nearly zero, demonstrating near-perfect color and structural fidelity.
• Demonstrated high robustness and generalization across different scenes and noise conditions.

Conclusions:

• The SG-ATV framework enables stable and high-quality passive NLOS imaging.
• The proposed prior-guided mechanism offers a scene-independent weighting strategy.
• This approach can be integrated into other NLOS reconstruction pipelines for broader applicability without additional parameter tuning.