Learning Dense Correspondences for Non-Rigid Point Clouds With Two-Stage Regression.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Geometry Processing

Background:

• Establishing dense correspondences between 3D shapes is crucial for tasks like registration and analysis.
• Partial and non-rigidly deformable point clouds present significant challenges due to missing data and complex shape variations.

Purpose of the Study:

• To develop a novel deep learning method for predicting dense correspondences from partial point clouds of non-rigidly deformable targets.
• To accurately estimate vertex displacements of a template mesh towards target point clouds.

Main Methods:

• A two-stage deep learning regression framework utilizing graph convolutional networks (GCNs) and attention mechanisms.
• Global regression network for initial coarse displacements using hierarchical encoder-decoder architecture.
• Local regression network for refining displacements by fusing local point cloud features and mesh graph features via attention.

Main Results:

• The proposed method accurately predicts dense correspondences for partial point clouds of deformable objects.
• Demonstrated robustness and accuracy across diverse datasets including human bodies, animals, and hands.
• Effective generalization to unseen real-world data through a robust fine-tuning strategy.

Conclusions:

• The novel two-stage deep learning approach effectively addresses the challenge of dense correspondence prediction for non-rigid point clouds.
• The method offers a robust and accurate solution for 3D shape analysis and comparison of deformable objects.
• Future work could explore real-time applications and extensions to dynamic scenes.