Learning Mesh Representations via Binary Space Partitioning Tree Networks.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Machine Learning

Background:

• Polygonal meshes are fundamental in 3D graphics but underutilized in deep learning.
• Current neural generative models for 3D shapes rely on computationally expensive implicit functions and iso-surfacing.

Purpose of the Study:

• To develop a novel deep learning approach for 3D shape representation and generation using polygonal meshes.
• To overcome the limitations of existing methods by leveraging Binary Space Partitioning (BSP).

Main Methods:

• Introduced BSP-Net, a neural network that utilizes BSP for unsupervised convex decomposition of 3D shapes.
• The network learns to represent shapes using a BSP-tree, with learned weights defining planes and convex sets.
• Directly generates watertight, low-polygon polygonal meshes from inferred convex sets.

Main Results:

• BSP-Net achieves competitive reconstruction quality compared to state-of-the-art methods.
• The generated meshes are compact, watertight, and adept at representing sharp geometric features.
• Demonstrated significantly fewer primitives required for reconstruction.

Conclusions:

• BSP-Net offers an efficient and effective method for deep learning on polygonal meshes.
• The approach enables direct, unsupervised generation of high-quality 3D meshes.
• Explored variations for enhanced flexibility and generative capabilities.