A Hierarchical Prior Mining Approach for Non-Local Multi-View Stereo.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Computational Geometry

Background:

• Multi-view stereo (MVS) is crucial for 3D geometry recovery from 2D images.
• Low-textured regions present a significant challenge, degrading reconstruction quality.
• Existing MVS methods often rely on local information and single priors, limiting performance.

Purpose of the Study:

• To develop a novel framework for non-local multi-view stereo.
• To enhance 3D reconstruction robustness and completeness, particularly in challenging low-textured areas.
• To leverage multi-source priors for improved geometry recovery.

Main Methods:

• Proposed a Hierarchical Prior Mining (HPM) framework for non-local MVS.
• Developed HPM-MVS using non-local operations for precise initial hypotheses and a planar prior model.
• Introduced HPM-MVS++ incorporating structured image information and spatial geometric relationships into probabilistic graphical models, yielding novel multi-view matching costs.

Main Results:

• HPM-MVS provides precise initial hypotheses and improved planar prior modeling.
• HPM-MVS++ significantly enhances robustness in challenging scenarios and improves reconstruction completeness.
• Experimental validation on ETH3D and Tanks & Temples datasets demonstrates superior performance and generalization.

Conclusions:

• The HPM framework effectively captures non-local structural cues and multi-source priors for superior MVS.
• HPM-MVS and HPM-MVS++ offer significant improvements over existing methods for 3D reconstruction.
• The approach shows strong performance and generalization capabilities on benchmark datasets.