Unsupervised Deep Learning Model for Homography Estimation

Abstract

This paper proposes an unsupervised deep learning model for homography estimation, addressing limitations of traditional feature-based methods and supervised learning approaches. By leveraging reprojection error as the optimization objective, the model eliminates the need for labeled data while achieving precise homography estimation. The framework comprises a feature extraction module, a feature difference module, and a homography regression network. Extensive experiments on the MS-COCO and HPatches datasets demonstrate that the proposed model achieves a mean reprojection error (MRE) of 3.67 pixels and an accuracy (ACC) of 88.3%, closely approaching the performance of supervised methods like DeepHomography while significantly outperforming classical SIFT + RANSAC. The model's lightweight design ensures efficient inference, requiring only 12ms per estimation, making it suitable for real-time applications. Ablation studies validate the effectiveness of key components such as the feature difference module and regularization loss, highlighting their contributions to performance improvement. Compared to traditional methods, the proposed approach exhibits superior robustness under varying lighting conditions, viewpoint changes, and noise interference. Moreover, it removes the dependency on labeled data, reducing application costs and barriers. This unsupervised framework presents a practical and efficient solution for homography estimation and offers potential for broader applications in multi-view geometry and 3D reconstruction tasks. This is an example abstract. It describes the content and purpose of the paper succinctly. The abstract should be single-spaced and use Times New Roman 10 pt font.