CPDC-MFNet: conditional point diffusion completion network with Muti-scale Feedback Refine for 3D Terracotta...

Area of Science:

• Digital Heritage
• Computer Vision
• Cultural Heritage Preservation

Background:

• Terracotta Warriors suffer damage due to antiquity and excavation challenges.
• Point cloud completion is crucial for restoring these cultural relics.
• Existing methods often lack diversity in point cloud completion results.

Purpose of the Study:

• To develop a novel neural network for efficient and diverse point cloud completion of Terracotta Warrior fragments.
• To address the slow generation speed issue associated with diffusion models in point cloud reconstruction.

Main Methods:

• A new neural network architecture is proposed for Terracotta Warriors fragments completion.
• The model employs a reduced sampling strategy during the reverse diffusion stage for faster coarse result generation.
• A multi-scale refine network and Partition Attention Sampling are utilized for enhanced feature representation and refinement.

Main Results:

• The proposed model demonstrates competitive performance compared to existing methods on real and public datasets.
• Experiments validate the model's effectiveness in point cloud completion for cultural relic restoration.
• The approach achieves faster generation speeds while maintaining diverse and accurate output.

Conclusions:

• The developed neural network offers an effective solution for the point cloud completion of damaged Terracotta Warriors.
• The method successfully balances generation speed with the diversity and accuracy of reconstructed cultural relics.
• This work contributes to advancing digital heritage preservation techniques through innovative AI applications.