Co-Clustering on Bipartite Graphs for Robust Model Fitting.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Graph-based methods are common for model fitting but lose crucial data-model associations.
• Existing methods struggle with data contaminated by outliers and noise.

Purpose of the Study:

• To propose a novel model fitting method, co-clustering on bipartite graphs (CBG), for estimating multiple model instances.
• To address the loss of association information in traditional graph-based approaches.
• To enhance robustness against outliers and noise in data.

Main Methods:

• Reformulating model fitting as bipartite graph partitioning.
• Employing bipartite graph reduction to remove outliers and invalid hypotheses, enhancing graph reliability and reducing complexity.
• Utilizing a co-clustering algorithm for optimal bipartite graph partitioning to directly estimate model instances without post-processing.

Main Results:

• The CBG method effectively estimates multiple model instances from noisy data.
• Bipartite graph reduction successfully eliminates outliers and improves graph reliability.
• The co-clustering approach directly yields model instances, eliminating the need for post-processing.
• The proposed method demonstrates superior fitting performance compared to state-of-the-art methods.

Conclusions:

• The CBG method offers a robust and accurate solution for model fitting in challenging datasets.
• Fully leveraging the duality of data points and model hypotheses on bipartite graphs leads to significant performance gains.
• CBG provides a more reliable and computationally efficient alternative to existing graph-based model fitting techniques.