Parameterize Structure With Differentiable Template for 3D Shape Generation

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Modeling

Background:

• Current 3D shape generation methods often rely on complex networks and hierarchical annotations.
• These methods may overlook fine-grained details within object parts.
• A need exists for simpler, more effective structural representations for 3D shape generation.

Purpose of the Study:

• To propose a novel method for structural representation in 3D shape generation.
• To enable reconstruction and generation of editable 3D shapes with detailed part semantics.
• To develop a simplified yet effective approach for learning the latent space of 3D shapes.

Main Methods:

• Parameterizing shared structures within categories using a differentiable template and fixed-length parameters.
• Calculating cuboids from parameters to define concrete shapes.
• Utilizing three-view rendering boundaries of cuboids to capture internal details.
• Representing shapes with parameters and three-view details for Surface of Distance Function (SDF) calculation.

Main Results:

• Developed simple and effective networks for 3D shape reconstruction and generation.
• Successfully reconstructed diverse shapes from point clouds with complex details.
• Generated novel 3D shapes with smooth interpolation capabilities.
• Demonstrated superior performance in reconstruction, generation, and interpolation tasks.

Conclusions:

• The proposed method offers a simplified and effective approach to 3D shape representation.
• It enables the generation and reconstruction of detailed and editable 3D shapes.
• The technique shows significant advantages over existing methods in various 3D shape manipulation tasks.