Growing self-reconstruction maps

Area of Science:

• Computer Vision
• Computational Geometry
• Machine Learning

Background:

• Surface reconstruction from point clouds is crucial for 3D modeling.
• Existing methods often struggle with complex topologies and varying resolutions.

Purpose of the Study:

• To propose a novel surface reconstruction method using growing self-organizing maps (SOMs).
• To develop a method capable of generating accurate two-manifold mesh representations from unstructured point clouds.

Main Methods:

• The proposed method, growing self-reconstruction maps (GSRMs), extends growing neural gas (GNG).
• Key modifications include competitive Hebbian learning (CHL), vertex insertion, and edge removal mechanisms.
• The algorithm incorporates triangular faces for mesh generation.

Main Results:

• GSRMs successfully learn surface geometry and topology from point cloud data.
• The method generates triangular two-manifold meshes of target objects.
• Generated models accurately approximate object shapes, including concave regions, boundaries, and holes.
• The approach supports mesh generation at different resolutions.

Conclusions:

• GSRMs offer an effective approach for high-fidelity surface reconstruction.
• The method demonstrates robustness in handling complex shapes and topological features.
• This technique advances the creation of detailed 3D models from point cloud data.