Learning Implicit Functions for Dense 3D Shape Correspondence of Generic Objects.

Area of Science:

• Computer Vision
• 3D Shape Analysis
• Machine Learning

Background:

• Conventional implicit functions struggle with topology-varying shapes.
• Unsupervised learning of dense 3D shape correspondence remains a challenge.

Purpose of the Study:

• To develop an unsupervised method for learning dense 3D shape correspondence for topology-varying generic objects.
• To enable accurate semantic point mapping across dissimilar 3D shapes.

Main Methods:

• A novel implicit function generating probabilistic embeddings for 3D points.
• An inverse mapping function from embedding space to corresponding 3D points.
• Joint learning with uncertainty-aware loss functions and an encoder for shape latent codes.

Main Results:

• Successful demonstration of unsupervised 3D semantic correspondence.
• Effective application in unsupervised shape segmentation tasks.
• Inference provides confidence scores and corresponding semantic points.

Conclusions:

• The proposed method effectively addresses dense 3D shape correspondence for generic objects with varying topology.
• The probabilistic embedding and inverse mapping approach offers a robust solution for unsupervised semantic mapping.
• The technique shows promise for applications in 3D shape analysis and computer graphics.