The effectiveness of detector combinations

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Image matching and registration are crucial for various applications, including 3D reconstruction and augmented reality.
• The performance of these tasks heavily relies on the accuracy and robustness of local feature detectors.
• Evaluating detector combinations is essential for optimizing performance.

Purpose of the Study:

• To evaluate the performance improvement gained by combining different local feature detectors for image matching and registration.
• To introduce and assess a novel hybrid algorithm, MDSS, for accurate image matching.

Main Methods:

• Integration of five representative interest point and region detectors into a testing framework.
• Performance comparison using correspondence count, repeatability rate, and Reconstruction Similarity (RS).
• Development and evaluation of the MDSS (Multi-scale Difference of Gaussians and Spatially Segmented) hybrid algorithm.

Main Results:

• The combination of DoG extremum and MSCR detectors demonstrated superior performance compared to individual detectors and other combinations.
• The proposed MDSS algorithm achieved an average RS rate exceeding standard SIFT and GLOH by over 3.56%.
• MDSS also exhibited reduced computational time compared to SIFT and GLOH.

Conclusions:

• Combining specific local feature detectors, such as DoG extremum and MSCR, significantly enhances image matching and registration performance.
• The MDSS algorithm represents a promising advancement in accurate and efficient image matching.
• Further research into hybrid detector strategies can lead to more robust computer vision systems.