Sparse-to-Local-Dense Matching for Geometry-Guided Correspondence Estimation.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Machine Learning

Background:

• Establishing reliable correspondences between image views is crucial for many computer vision tasks.
• Existing methods often face challenges with computational efficiency and handling noise in 3D data.

Purpose of the Study:

• To propose a novel sparse-to-local-dense (S2LD) matching method for fully differentiable correspondence estimation.
• To improve accuracy and reduce computational cost in matching tasks.
• To enhance robustness against noise in 3D geometric estimation.

Main Methods:

• Developed a sparse-to-local-dense (S2LD) matching strategy leveraging epipolar geometry.
• Utilized an attention mechanism for view-conditioned feature description with a global receptive field.
• Introduced a 3D noise-aware regularizer with differentiable triangulation to handle supervision noise.

Main Results:

• Achieved outstanding matching accuracy across multiple datasets and tasks.
• Demonstrated superior geometric estimation capabilities.
• The S2LD method effectively reduces matching computation while maintaining sub-pixel accuracy.

Conclusions:

• The proposed S2LD method offers a robust and efficient solution for correspondence estimation.
• The integration of epipolar geometry and 3D noise awareness significantly improves performance.
• This approach advances the state-of-the-art in computer vision correspondence matching and geometric estimation.