Homography Estimation From Image Pairs With Hierarchical Convolutional Networks

Farzan Erlik Nowruzi, Robert Laganiere, Nathalie Japkowicz; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 913-920

In this paper, we introduce a hierarchy of twin convolutional regression networks to estimate the homography between a pair of images. In this framework, networks are stacked sequentially in order to reduce error bounds of the estimate. At every convolutional network module, features from each image are extracted independently, given a shared set of kernels, also known as Siamese network model. Later on in the process, they are merged together to estimate the homography. Further, we evaluate and compare effects of various training parameters in this context. We show that given the iterative nature of the framework, highly complicated models are not necessarily required, and high performance is achieved via hierarchical arrangement of simple models. Effectiveness of the proposed method is shown through experiments on MSCOCO dataset, in which it significantly outperforms the state-of-the-art.