Exploiting Visibility Information in Surface Reconstruction to Preserve Weakly Supported Surfaces

Area of Science:

• Computer Vision
• Computational Geometry
• 3D Reconstruction

Background:

• Traditional 3D surface reconstruction methods often struggle with surfaces lacking sufficient input points.
• Visibility information from 3D points is underexplored for surface reconstruction.

Purpose of the Study:

• To develop a novel 3D surface reconstruction method that overcomes limitations of existing approaches.
• To reconstruct surfaces even when they do not contain input points, by leveraging visibility data.

Main Methods:

• Modeling free space using visibility information of input 3D points.
• Introducing a novel interface classifier to detect surfaces between free and full space.
• Modifying a state-of-the-art reconstruction method with the proposed interface classifier.

Main Results:

• The method successfully reconstructs surfaces not containing input points, revealing occluders through occlusion.
• The enhanced method demonstrates improved accuracy and ability to reconstruct weakly supported surfaces.
• Performance is validated on datasets with noise, undersampling, and outliers.

Conclusions:

• The proposed free space modeling and interface classification approach offers a robust solution for 3D surface reconstruction.
• This method advances the capability to reconstruct complex and sparsely sampled surfaces.
• Visibility information is a powerful cue for inferring object geometry and occlusions.