Normal Reorientation for Scene Consistency.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Processing

Background:

• 3D scanning technology has advanced, leading to more captured indoor scenes.
• Generating orientation-consistent normals for indoor point clouds is crucial but challenging.
• Existing methods focus on object-level targets, struggling with real-scanned indoor data.

Purpose of the Study:

• To address the difficulty of computing consistent surface orientation for real scanned indoor point clouds.
• To propose a novel normal reorienting framework for indoor scene consistency.
• To achieve globally consistent normal orientations in indoor point cloud data.

Main Methods:

• Estimating initial normals and extracting connected regions from indoor point clouds.
• Designing an abstract orientation bridging tree (OBT) for hierarchical region organization.
• Implementing iterative orientation propagations and a voting mechanism for parent-child region consistency.

Main Results:

• The proposed NRSC framework successfully generates locally orientation-consistent regions.
• A voting mechanism effectively rectifies child node orientations based on parent nodes.
• Experimental results demonstrate globally consistent orientation generation for indoor point clouds.

Conclusions:

• The NRSC framework effectively overcomes challenges in orienting normals for indoor point clouds.
• The OBT and voting mechanism are key components for achieving scene-wide orientation consistency.
• The method shows robust performance on both fused and single-view indoor scenes.