Deep Homography Estimation based on Attention Mechanism

Abstract

Since the existing supervised learning has a strong dependence on the real ground labeling and ignores the importance of depth differences and moving objects in the image, an unsupervised homography estimation algorithm is proposed. Firstly, a resnet34 backbone network is constructed, and two feature extraction modules with shared weights are used. Then, each initial feature extraction module is embedded with an Shuffle attention mechanism (SA), which is used to extract features that can provide greater help for model training. Secondly, the triple loss function is used as the loss function of the neural network, so that the neural network can better learn the difference of the input image, and learn the alignment between the corrected image generated by the estimated homography matrix and the input image. Finally, the proposed algorithm is compared with the existing methods, and the results show that the proposed method has good adaptability and performance in low texture and low brightness images.