A Bayesian hierarchical factorization model for vector fields.

Area of Science:

• Data Science
• Computer Vision
• Machine Learning

Background:

• Factorization methods are crucial for multi-dimensional data analysis, but often handle only scalar values.
• Existing models lack the ability to process complex observations like vector fields or account for spatial dependencies.

Purpose of the Study:

• To propose a probabilistic factorization model capable of handling vector arrays and non-exchangeable factors.
• To incorporate spatial relationships within the factorization process for enhanced data analysis.

Main Methods:

• Developed a Bayesian hierarchical model treating factors as latent random variables.
• Introduced a Markov structure to capture spatial dependencies between data dimensions.
• Designed a specialized observation model for vector arrays from continuous domains.

Main Results:

• The proposed model effectively represents vector arrays from continuous fields.
• Demonstrated success in analyzing optical flow fields from synthetic and real-world image data.

Conclusions:

• The novel probabilistic factorization model offers a powerful tool for analyzing complex spatial data.
• The incorporation of spatial structures and vector observations advances factorization techniques for computer vision applications.