High accuracy and visibility-consistent dense multiview stereo.

Area of Science:

• Computer Vision
• Photogrammetry
• 3D Reconstruction

Background:

• Dense multiview stereo (MVS) methods have advanced significantly but struggle with uncontrolled outdoor environments.
• Existing MVS techniques often fail to accurately reconstruct large-scale scenes under variable lighting and conditions.

Purpose of the Study:

• To develop a robust and accurate dense multiview stereo pipeline capable of handling large-scale outdoor scenes.
• To overcome limitations of current MVS methods in uncontrolled imaging conditions.

Main Methods:

• A novel pipeline combining minimum s-t cut optimization on an adaptive domain for outlier filtering and initial surface reconstruction.
• Integration of visibility constraints and mesh-based variational refinement for detailed surface reconstruction and handling of photo-consistency and regularization.
• Adaptive resolution strategies to manage computational complexity and detail capture.

Main Results:

• The pipeline successfully reconstructs highly detailed 3D models of diverse scenes, including compact objects, architectural sites, landscapes, and cultural heritage.
• Demonstrated superior accuracy on the Strecha et al. dense multiview benchmark compared to state-of-the-art methods.
• Achieved reconstructions within reasonable timeframes.

Conclusions:

• The proposed dense multiview stereo pipeline offers a significant advancement for 3D reconstruction in challenging outdoor environments.
• The method provides a robust and accurate solution for large-scale scene reconstruction without sacrificing detail or efficiency.