Indoor Scene Reconstruction Using Near-Light Photometric Stereo

Area of Science:

• Computer Vision
• Photogrammetry
• 3D Reconstruction

Background:

• Photometric stereo (PS) traditionally requires controlled lighting.
• Low-light conditions and uncalibrated illumination pose significant challenges for PS.
• Existing methods struggle with albedo variations, shadowing, and noise.

Purpose of the Study:

• To develop a novel framework for photometric stereo (PS) under challenging low-light conditions.
• To enable 3D surface reconstruction using uncalibrated, near-light illumination from free-form video sequences.
• To address and overcome limitations like albedo variations, shadowing, perspective projections, and camera noise.

Main Methods:

• Utilizes specular spheres for robust light position triangulation via a perspective-correcting Hough transform and least-squares approach.
• Employs an iterative reweighting scheme with an ℓp-norm minimizer for robustly solving the calibrated near-light PS problem.
• Operates on minimalistic, affordable setups capturing free-form video sequences.

Main Results:

• Simultaneously reconstructs depth, albedo (relative to light source intensity), and surface normals.
• Demonstrates robust performance on both synthetic and real-world scenes.
• Effectively handles issues such as albedo variations, shadowing, perspective projections, and camera noise.

Conclusions:

• The proposed framework offers a robust and versatile solution for photometric stereo in low-light, uncalibrated environments.
• It provides simultaneous 3D reconstruction of geometry and surface properties from affordable video setups.
• This advancement expands the applicability of PS to previously challenging scenarios.