End-to-End Matching Network for Invariant Local Features

The accurate estimation of the pose is the critical step of visual localization. Aiming at the problem that the current feature tasks cannot cope well with the end-to-end framework of scene transformation and feature extraction, an end-to-end matching network integrating feature point extraction, descriptor construction, and local feature matching is proposed. The feature points extraction based on the neural network is used with descriptor construction to form a joint training network, which obtains local features with a robust viewpoint and illumination changes. Introduce Attentional Graph Neural Networks (AGNN) to enhance the connection of local features of image pairs. The experimental results show that the proposed method can achieve end-to-end local feature matching tasks and meet the requirements of the front-end of the visual localization system for environmental resilience. Compared with classical algorithms, homography estimation, matching precision, and recall have a better performance.