Deep photometric stereo for dynamic single-pixel 3D imaging.

Area of Science:

• Computer Vision
• Optical Imaging
• 3D Reconstruction

Background:

• Dynamic single-pixel 3D imaging (SPI) faces challenges with calibration complexity and resolution-measurement trade-offs.
• Existing methods often require intricate calibration procedures, limiting their practical application.

Purpose of the Study:

• To develop a calibration-free framework for dynamic 3D SPI video.
• To achieve high-resolution and accurate 3D surface normal recovery without prior system calibration.

Main Methods:

• Integration of binocular single-pixel imaging (SPI) with a super-resolution photometric stereo network (SRPS-Net).
• Reconstruction of photometric images from arbitrary viewpoints, followed by SRPS-Net processing for surface normal estimation.
• A novel framework that bypasses the need for traditional system calibration.

Main Results:

• Successful dynamic 3D SPI video reconstruction at 128x128 resolution and 6.5 fps.
• Achieved pixel-level accuracy in surface normal recovery.
• Demonstrated robust generalization across complex objects and dynamic gestures.

Conclusions:

• The proposed method offers a compact, cost-effective, and calibration-free solution for dynamic single-pixel 3D imaging.
• This framework significantly advances the capabilities of real-time 3D imaging systems.
• Enables new applications in fields requiring dynamic, high-fidelity 3D scene capture.