A Concurrent Multiscale Detector for End-to-End Image Matching.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• End-to-end image matching is crucial for various computer vision tasks.
• Existing deep matching networks face challenges in detecting repeatable and discriminative key points.

Purpose of the Study:

• To improve deep matching networks for enhanced key-point detection and descriptor extraction.
• To enhance the accuracy and repeatability of image matching.

Main Methods:

• Proposing a concurrent multiscale detector (CS-det) network with parallel convolutional networks for multiscale feature extraction.
• Integrating an attention module for adaptive fusion of feature response maps.
• Introducing two novel rank consistent losses (RC-losses): score rank consistent loss (RC-S-loss) for repeatability and score-discrimination RC-loss for discrimination.

Main Results:

• The CS-det network improved the mean matching results of deep detectors by 1.4%-2.1%.
• The proposed RC-losses boosted matching performances by 2.7%-3.4% compared to score difference loss.
• Demonstrated significant improvements in key-point repeatability and discrimination.

Conclusions:

• The proposed CS-det network and RC-losses effectively enhance end-to-end image matching performance.
• The novel loss functions address limitations of traditional methods by focusing on relative key-point scores.
• The advancements contribute to more accurate and robust image matching in computer vision.