Learning Geometric Feature Embedding with Transformers for Image Matching.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Local feature matching is fundamental to many computer vision applications, including camera pose estimation.
• Existing methods often struggle to effectively incorporate geometric information, which is crucial for robust matching.

Purpose of the Study:

• To propose a novel geometric feature embedding matching (GFM) method for enhanced local feature matching.
• To improve the accuracy and robustness of feature matching by effectively utilizing geometric information.

Main Methods:

• Introduced an adaptive keypoint geometric embedding module for dynamic adjustment of keypoint position.
• Implemented an orientation geometric embedding module for modeling rotational geometric information.
• Utilized interleaved self-attention and cross-attention mechanisms for local feature enhancement.
• Employed dual-softmax for solving correspondences after multiplying with local features.

Main Results:

• The proposed GFM method demonstrated satisfactory performance across three diverse datasets: MegaDepth, Hpatches, and Aachen Day-Night v1.1.
• The method effectively integrates geometric information, leading to improved feature matching results.
• Validation confirmed the effectiveness of the adaptive keypoint and orientation embedding modules.

Conclusions:

• The geometric feature embedding matching (GFM) method offers a significant advancement in local feature matching.
• GFM's ability to dynamically incorporate geometric and orientation information enhances performance in challenging visual scenes.
• The method provides a robust solution for downstream tasks requiring accurate local feature correspondences.