LogDet Rank Minimization with Application to Subspace Clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Matrix Analysis

Background:

• Low-rank matrix approximation is crucial for machine learning and computer vision tasks.
• The nuclear norm is a common convex surrogate for rank, but can poorly approximate the true rank.
• Existing methods struggle with accurate rank approximation in practical scenarios.

Purpose of the Study:

• To propose a novel log-determinant (LogDet) function as a smoother and closer approximation to the rank operator.
• To develop a low-rank representation method for subspace clustering using the LogDet function.
• To enhance the performance of subspace clustering algorithms.

Main Methods:

• Utilizing a log-determinant (LogDet) function as a nonconvex approximation to the rank.
• Applying the Augmented Lagrange multipliers strategy for iterative optimization of the LogDet objective function.
• Constructing an affinity graph matrix using angular information of principal directions for spectral clustering.

Main Results:

• The proposed LogDet-based method provides a more accurate low-rank representation compared to nuclear norm methods.
• The approach effectively handles potentially large-scale data through iterative optimization.
• Experimental results demonstrate superior performance in motion segmentation and face clustering tasks over state-of-the-art algorithms.

Conclusions:

• The LogDet function offers a promising alternative to the nuclear norm for low-rank approximation in subspace clustering.
• The proposed method achieves state-of-the-art results in benchmark datasets.
• This work advances low-rank representation techniques for computer vision and machine learning applications.