PRS-Net: Planar Reflective Symmetry Detection Net for 3D Models.

Area of Science:

• Computer Graphics
• Computational Geometry
• Artificial Intelligence

Background:

• Symmetry is crucial for understanding 3D models, aiding tasks like segmentation and matching.
• Existing methods for detecting planar reflective symmetry are often slow and may miss some symmetry planes.
• Automated symmetry detection is vital for efficient 3D shape analysis.

Purpose of the Study:

• To develop a novel, efficient, and accurate framework for discovering global planar reflective symmetry in 3D shapes.
• To overcome the limitations of traditional sampling-based methods in terms of speed and completeness.
• To enable robust symmetry detection even with incomplete or noisy 3D model data.

Main Methods:

• An unsupervised 3D convolutional neural network (CNN) framework is proposed for symmetry detection.
• Input 3D shapes are represented as voxels for feature extraction by the CNN.
• A specialized symmetry distance loss and regularization loss are introduced to refine symmetry plane identification and prevent duplication.

Main Results:

• The framework successfully identifies global planar reflective symmetry and rotation axes for generalized cylinders.
• A post-processing step effectively removes invalid and duplicate symmetry planes and axes.
• The neural network-based approach is significantly faster (hundreds of times) than state-of-the-art sampling methods.

Conclusions:

• The proposed learning framework provides a reliable and accurate method for automated 3D shape symmetry discovery.
• The approach demonstrates superior speed and robustness compared to traditional techniques.
• This method has broad implications for various geometry processing applications requiring structural understanding of 3D models.