OANet: Learning Two-View Correspondences and Geometry Using Order-Aware Network.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Accurate image correspondence is crucial for computer vision tasks.
• Existing methods often struggle to effectively integrate local and global spatial context.

Purpose of the Study:

• To propose a novel hierarchical network, Order-Aware Network (OANet), for inferring correspondence probabilities and relative pose.
• To improve the accuracy of two-view geometry estimation and feature matching.

Main Methods:

• A hierarchical network architecture is introduced, comprising three key operations.
• Soft assignment clustering to capture local context and create permutation-invariant ordered clusters.
• Spatial correlation of clusters to encode global context, followed by interpolation for hierarchical integration.

Main Results:

• Significantly improved accuracy in two-view geometry and correspondence estimation on outdoor and indoor datasets.
• Achieved first place in the CVPR 2019 image matching workshop challenge.
• Attained state-of-the-art results on the Visual Localization benchmark.

Conclusions:

• The proposed Order-Aware Network effectively integrates local and global spatial context for robust image correspondence.
• The method demonstrates superior performance in geometric accuracy and visual localization tasks.
• The hierarchical approach offers a powerful framework for feature matching and pose regression.