Geometry-Aware Hierarchical Bayesian Learning on Manifolds.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Bayesian learning with Gaussian processes excels in computer vision but is understudied for 3D manifold data like meshes and point clouds.
• A key challenge is efficiently aggregating geometric features from irregular 3D inputs.

Purpose of the Study:

• To propose a hierarchical Bayesian learning model for effective geometric feature aggregation on 3D manifold-valued vision data.
• To enable geometrically aware inferences on manifolds without manual feature descriptors.
• To explore joint learning by integrating neural networks with Bayesian models.

Main Methods:

• Developed a novel geometry-aware kernel with intra-kernel convolution for manifold data.
• Employed Gaussian process regression for input organization.
• Implemented a hierarchical Bayesian network for feature aggregation.
• Integrated neural network feature learning with Bayesian model feature aggregation.

Main Results:

• The proposed hierarchical Bayesian model outperforms existing Bayesian methods on manifold data.
• Demonstrated the effectiveness of geometry-aware kernels for irregular 3D inputs.
• Showcased the potential of coupling neural networks with Bayesian networks for joint learning on manifolds.

Conclusions:

• The hierarchical Bayesian model offers a robust solution for feature aggregation in 3D manifold vision.
• The integration of geometry-aware kernels and hybrid neural-Bayesian approaches advances manifold learning.
• This work opens new avenues for applying Bayesian methods to complex 3D vision problems.