NCMNet: Neighbor Consistency Mining Network for Two-View Correspondence Pruning.

Area of Science:

• Computer Vision
• Geometric Deep Learning

Background:

• Correspondence pruning is vital for feature matching tasks, aiming to identify correct correspondences (inliers) from initial sets.
• Existing methods struggle with outliers due to similarity constraints, misclassifying them as neighbors.
• The presence of numerous false correspondences (outliers) near inliers complicates accurate neighbor identification.

Purpose of the Study:

• To propose a novel global-graph space for identifying consistent neighbors based on graph structures.
• To enhance the robustness of correspondence pruning for diverse matching scenarios.
• To introduce the Neighbor Consistency Mining Network (NCMNet) for outlier removal and model estimation.

Main Methods:

• Utilizing a global connected graph to represent affinity relationships between correspondences based on spatial and feature consistency.
• Developing a neighbor consistency block to leverage three types of neighbors by extracting intra-neighbor context and exploring inter-neighbor interactions.
• Implementing NCMNet to progressively mine neighbor consistency for accurate outlier removal.

Main Results:

• NCMNet significantly outperforms state-of-the-art methods in two-view geometry estimation benchmarks.
• The method demonstrates strong generalization capabilities across various extended tasks.
• Experimental results validate the effectiveness of the proposed global-graph space and neighbor consistency block.

Conclusions:

• The proposed global-graph space and neighbor consistency mining effectively address the limitations of traditional nearest neighbor strategies.
• NCMNet offers a robust and accurate solution for correspondence pruning in computer vision.
• The method shows significant potential for applications in remote sensing, 3D reconstruction, and visual localization.