U-Match: Exploring Hierarchy-Aware Local Context for Two-View Correspondence Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Feature-based two-view geometry estimation critically relies on accurate outlier rejection.
• Current methods struggle with severe outliers and uncertain inlier/outlier distributions, limiting local context extraction.
• Explicit neighborhood modeling at fixed scales lacks flexibility for complex correspondence scenarios.

Purpose of the Study:

• To propose a novel network, U-Match, for flexible implicit local context awareness at multiple levels.
• To address the limitations of existing methods in handling outliers and extracting reliable local context.
• To improve the accuracy and robustness of feature-based geometry estimation.

Main Methods:

• Designed a hierarchy-aware graph representation module for implicit multi-level local context aggregation.
• Introduced an orthogonal local-and-global information fusion module to integrate complementary contexts efficiently.
• Developed a novel network architecture enabling flexible and implicit context awareness.

Main Results:

• U-Match demonstrates remarkable capabilities in rejecting outlier correspondences.
• The network effectively aggregates multi-level local context and fuses it with global context.
• Achieved state-of-the-art performance across various computer vision tasks.

Conclusions:

• U-Match offers a flexible and effective solution for feature matching and geometry estimation.
• Implicit multi-level context awareness circumvents limitations of explicit neighborhood modeling.
• The proposed method significantly advances the field of two-view geometry estimation.