PSDF: Prior-Driven Neural Implicit Surface Learning for Multi-View Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Deep Learning

Background:

• Traditional Multi-View Stereo (MVS) pipelines struggle with noisy and incomplete surface geometry, particularly in low-texture or unevenly lit areas.
• Neural Implicit Surface Reconstruction (NISR) offers a promising alternative but faces challenges in accurately reconstructing delicate geometry from uncontrolled real-world scenes due to underconstrained optimization.

Purpose of the Study:

• To develop a novel framework (PSDF) that enhances neural implicit surface reconstruction quality.
• To leverage both external geometric priors from pre-trained MVS networks and internal priors from NISR models.
• To improve the accuracy and fidelity of reconstructed surfaces, especially in complex, uncontrolled environments.

Main Methods:

• Introduced a framework (PSDF) combining external geometric priors from MVS and internal priors from NISR.
• Implemented a visibility-aware feature consistency loss and depth prior-assisted sampling using external MVS priors.
• Utilized internal prior-guided importance rendering to mitigate biased rendering issues inherent in NISR.

Main Results:

• The proposed methods provide strong geometric consistency constraints and improve surface intersection point localization.
• PSDF significantly enhances the accuracy and detail of neural implicit surface reconstruction.
• Experiments on the Tanks and Temples datasets demonstrate state-of-the-art performance on complex, uncontrolled scenes.

Conclusions:

• The PSDF framework effectively addresses the limitations of traditional MVS and basic NISR for surface reconstruction.
• Integrating external and internal geometric priors is crucial for high-quality neural implicit surface learning.
• PSDF achieves superior results in reconstructing delicate geometry from challenging real-world data.