Learning Two-View Correspondences and Geometry via Local Neighborhood Correlation

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Feature correspondences are crucial for computer vision tasks like camera pose estimation.
• Existing methods struggle with outlier rejection in complex scenes.

Purpose of the Study:

• To propose a novel network, LNCNet, for robust feature matching and outlier rejection.
• To enhance camera pose estimation accuracy using local contextual information.

Main Methods:

• Utilized the k-Nearest Neighbor (KNN) algorithm for initial local neighborhood division.
• Developed a Local Neighborhood Correlation (LNC) matrix to filter outliers.
• Clustered filtered information into feature vectors for inlier probability determination.

Main Results:

• LNCNet effectively captures contextual information within local regions.
• The proposed method demonstrates superior performance in outlier rejection compared to state-of-the-art networks.
• Achieved improved camera pose estimation accuracy in diverse indoor and outdoor scenes.

Conclusions:

• LNCNet offers a stable local constraint for accurate feature correspondence.
• The network provides a robust solution for challenging computer vision problems.
• LNCNet advances the state-of-the-art in outlier rejection and camera pose estimation.