Homography Estimation Network Based on Dense Correspondence

Abstract

Reliable homography estimation is essential for plenty of computer vision tasks. However, the ground truth of homography is difficult to obtain. Moreover, it is relatively hard for ordinary convolutional neural networks to model correspondence, which leads to the difficulty for optimization. In this paper, we propose a novel framework based on unsupervised dense correspondence. We extract feature pyramid for each image and estimate the dense correspondence map in a coarse-to-fine manner. Relying on correlation layers and photometric loss, we can get rid of the deficiency of ordinary networks and estimate reliable dense correspondence for each image pair, which facilitates the later prediction of homography matrix. To achieve the homography training on both synthetic dataset and realistic dataset, we apply different losses for different tasks respectively being supervised loss and triplet loss, to regulate the training procedure. Qualitative and quantitative evaluations illustrate the effectiveness of the dense correspondence map, and shows the competitive performance of our method compared with traditional feature-based algorithms and deep learning methods.

This research is partly supported by NSFC, China (No: 61876107, U1803261).