Efficient Dynamic Correspondence Network.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Dense correspondence matching is crucial for image analysis but computationally expensive.
• Existing methods often use inefficient 4D convolutions and fixed feature learning, limiting robustness.

Purpose of the Study:

• To develop an efficient and robust method for establishing dense correspondences between image pairs.
• To address the limitations of quadratic complexity and fixed feature learning in current dense matching techniques.

Main Methods:

• Propose the Efficient Dynamic Correspondence Network (EDCNet) using a coarse-to-fine architecture.
• Incorporate a neighborhood aggregation (NA) module with linear complexity and a dynamic feature learning (DFL) module for flexibility.
• Utilize pre-separate convolution (Psconv) and dynamic convolution (Dyconv) for efficient and adaptable feature extraction.

Main Results:

• The proposed EDCNet achieves efficient and effective dense correspondence establishment.
• The NA module demonstrates high efficiency with linear complexity.
• The DFL module provides enhanced robustness to various challenging scenarios.
• Experiments show favorable performance against state-of-the-art methods on HPatches, Aachen Day-Night, and InLoc datasets.

Conclusions:

• EDCNet offers an efficient and robust solution for dense image correspondence.
• The combination of NA and DFL modules enhances performance and adaptability.
• The method shows significant improvements over existing approaches in challenging real-world scenarios.