DualRC: A Dual-Resolution Learning Framework With Neighbourhood Consensus for Visual Correspondences

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Establishing accurate correspondences between images is crucial for various computer vision tasks.
• Existing methods often struggle with both geometric and semantic matching or lack adaptability.

Purpose of the Study:

• To develop a flexible and accurate framework for image correspondence.
• To improve both geometric and semantic matching capabilities.
• To offer adaptable model variants for specific matching needs.

Main Methods:

• An end-to-end trainable framework employing a coarse-to-fine matching strategy.
• Utilizing multi-resolution feature maps and 4D convolutions for neighborhood consensus.
• Introducing three model variants: DualRC (universal), DualRC-L (efficient geometric), and DualRC-D (dynamic consensus).

Main Results:

• Demonstrated superior performance in both geometric and semantic matching tasks on public benchmarks.
• DualRC-L showed significant acceleration for high-resolution images.
• DualRC-D effectively handled scale variations through dynamic neighborhood consensus.

Conclusions:

• The proposed framework provides a robust and adaptable solution for image correspondence.
• The specialized variants offer tailored performance for different matching scenarios.
• This work advances the state-of-the-art in image matching techniques.