Isomorphic Mesh Generation From Point Clouds With Multilayer Perceptrons.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• Generating consistent 3D mesh structures from point clouds is challenging due to noise and missing data.
• Existing methods often require extensive pre-processing or class-specific models.
• Unified mesh representations are needed for efficient deep neural network (DNN) integration.

Purpose of the Study:

• To introduce a novel neural network, the isomorphic mesh generator (iMG).
• To enable the generation of isomorphic meshes from noisy and incomplete point clouds.
• To facilitate efficient processing of 3D surface models by DNNs.

Main Methods:

• The iMG employs a data-free approach, not requiring pre-existing training datasets.
• It utilizes a step-by-step mapping strategy to deform a reference mesh onto the input point cloud.
• This strategy ensures stable mapping, flexible deformation, and preservation of the reference mesh structure.

Main Results:

• The iMG successfully generates isomorphic meshes from point clouds with noise and missing parts.
• Isomorphic meshes provide a unified structure applicable across different object classes.
• The method demonstrated reliable performance in simulations and experiments on a mobile phone.

Conclusions:

• The iMG offers a robust solution for generating consistent 3D mesh representations.
• Isomorphic meshes streamline DNN applications by eliminating the need for class-specific pre-processing.
• This approach enhances computational efficiency in memory usage and calculation time for 3D data processing.