Exploiting Unsupervised and Supervised Constraints for Subspace Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Data Mining

Background:

• Image and video analysis often involves data from multiple low-dimensional subspaces, each representing a class.
• Subspace clustering aims to separate data points based on these underlying subspaces.
• Existing methods typically use unconstrained subspace models, which may not accurately represent real-world data.

Purpose of the Study:

• To develop a subspace clustering method based on a constrained subspace assumption, better reflecting real-world data properties.
• To introduce a unified optimization framework capable of incorporating various forms of supervised information.
• To enhance clustering accuracy in image and video analysis tasks.

Main Methods:

• Proposing a constrained subspace assumption, where data points are restricted within their corresponding subspaces (e.g., submanifolds, spatial regularity).
• Utilizing a unified integer linear programming optimization framework.
• Employing a branch-and-bound (BB) method for efficient solving of the optimization problem.
• Integrating supervised information such as subspace number, outlier ratio, and pairwise constraints.

Main Results:

• The proposed constrained subspace clustering method significantly outperforms state-of-the-art algorithms in clustering accuracy on real-world data.
• Demonstrated the effectiveness of the framework in incorporating and leveraging supervised information.
• The method shows improved performance in scenarios with structured data, like moving objects in videos and faces under varying illumination.

Conclusions:

• Constrained subspace clustering provides a more accurate model for real-world image and video data.
• The proposed integer linear programming framework offers a flexible and powerful approach to subspace clustering.
• Incorporating supervised information further boosts the performance and applicability of the clustering method.