GenPoly: Learning Generalized and Tessellated Shape Priors via 3D Polymorphic Evolving.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Existing 3D prior models often sacrifice fine geometric details by focusing on coarse object decomposition.
• Generalizable 3D representations are crucial for diverse generation tasks but struggle with intricate local variations.

Purpose of the Study:

• To introduce GenPoly, a novel generalized 3D prior model for multiple 3D generation tasks.
• To develop a method that preserves and incorporates fine local geometric details for high-fidelity 3D shape generation.

Main Methods:

• Formulating 3D prior modeling as a bottom-up polymorphic evolving process.
• Introducing a polymorphic variational autoencoder (PolyVAE) with a versatile shape residual codebook for encoding local geometry.
• Developing a 3D polymorphic evolving scheme for progressive, coarse-to-fine local detail refinement.

Main Results:

• GenPoly effectively preserves intricate local geometry, overcoming limitations of previous coarse-grained approaches.
• The proposed PolyVAE and evolving scheme enable the generation of visually compelling 3D shapes with rich details.
• GenPoly demonstrates superior performance over state-of-the-art methods in downstream tasks, especially in local detail preservation.

Conclusions:

• GenPoly offers a new paradigm for 3D prior modeling by leveraging bottom-up local geometry refinement.
• The method achieves high-fidelity 3D generation with superior preservation of fine details.
• GenPoly represents a significant advancement for detailed and generalizable 3D shape synthesis.