Learning Feature Matching via Matchable Keypoint-Assisted Graph Neural Network.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Local feature matching is crucial for 3D scene reconstruction.
• Existing attention-based GNNs struggle with non-repeatable keypoints, impacting efficiency and accuracy.
• Redundant connections in fully-connected graphs hinder performance.

Purpose of the Study:

• To develop a more efficient and accurate GNN for local feature matching.
• To address the limitations of non-repeatable keypoints in graph-based methods.
• To propose a sparse attention-based GNN architecture.

Main Methods:

• Introduced MaKeGNN, a sparse attention-based GNN.
• Developed Bilateral Context-Aware Sampling (BCAS) to select reliable keypoints.
• Utilized Matchable Keypoint-Assisted Context Aggregation (MKACA) for focused message passing.

Main Results:

• MaKeGNN significantly outperforms state-of-the-art methods on challenging benchmarks.
• Achieved superior accuracy in local feature matching.
• Demonstrated substantial reductions in computational and memory complexity.

Conclusions:

• MaKeGNN offers a superior balance of accuracy and efficiency for local feature matching.
• The proposed BCAS and MKACA modules effectively handle non-repeatable keypoints.
• This approach advances the field of computer vision for 3D scene understanding.