DeMatch++: Two-View Correspondence Learning via Deep Motion Field Decomposition and Respective Local-Context...

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Two-view correspondence learning prioritizes motion field coherence and smoothness.
• Conventional methods struggle with computational cost or large scene disparities.

Purpose of the Study:

• Introduce DeMatch++, a novel network for improved motion field decomposition.
• Achieve implicit regularization with lower computational overhead and inherent piecewise smoothness.

Main Methods:

• Decompose motion fields into low-frequency components using Fourier principles.
• Utilize linearly independent basis vectors for smooth sub-field generation.
• Aggregate local and global context, employing masked decomposition and compact representation.

Main Results:

• DeMatch++ outperforms state-of-the-art methods in motion field recovery.
• Demonstrates computational efficiency and piecewise smoothness.
• Successfully recovers cleaner motion fields for precise vector derivation.

Conclusions:

• DeMatch++ offers a robust and efficient solution for two-view correspondence.
• The discrete learnable architecture avoids dense field computation.
• Publicly available code and models facilitate further research.