Common Visual Pattern Discovery via Nonlinear Mean Shift Clustering.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Discovering common visual patterns (CVPs) between images is difficult due to geometric/photometric changes, noise, and clutter.
• Traditional methods rely on feature correspondence and graph optimization, which are computationally expensive.

Purpose of the Study:

• To propose an efficient and robust approach for discovering common visual patterns (CVPs) in images.
• To overcome the high computational cost associated with conventional methods.

Main Methods:

• Constructing a geometric space using a matrix Lie group structure from estimated transformations of local interest regions.
• Employing mean shift clustering in the Lie algebra vector space to group similar transformations.
• Utilizing a novel distance measure for non-Euclidean data within the matrix Lie group.

Main Results:

• Successfully identifies common visual patterns (CVPs) by grouping coherently deformed local regions.
• Demonstrates robustness and efficiency in single and multiple common object discovery tasks.
• Validates performance on near-duplicate image retrieval tasks.

Conclusions:

• The proposed method offers an efficient alternative to traditional graph-based optimization for CVP discovery.
• The novel perspective utilizing matrix Lie groups and mean shift clustering enhances performance and reduces computational load.