Robust Feature Matching Using Spatial Clustering with Heavy Outliers.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional feature matching relies on descriptor similarity, often requiring predefined transformation models.
• Existing methods struggle with complex or varying image transformations, limiting their real-world applicability.
• Predefined models are insufficient for diverse and challenging real-world scenarios.

Purpose of the Study:

• To develop a novel, robust method for removing mismatches in putative feature matches.
• To address the limitations of existing approaches that require predefined geometric constraints.
• To improve the accuracy and applicability of feature matching across various image transformations.

Main Methods:

• The study frames feature matching as a spatial clustering problem with outliers.
• A customized Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is adapted for feature matching.
• An iterative clustering strategy is employed to enhance performance with degraded data.

Main Results:

• The proposed method achieves quasi-linear time complexity through customized DBSCAN.
• Experiments demonstrate superior performance over state-of-the-art alternatives on diverse datasets.
• The approach shows promising results in near-duplicate image retrieval and co-segmentation tasks.

Conclusions:

• The spatial clustering approach offers a more flexible and effective solution for feature matching mismatch removal.
• The customized DBSCAN method provides a robust and efficient alternative to model-based techniques.
• The method's adaptability makes it suitable for challenging real-world computer vision applications.