Rethinking Planar Homography Estimation Using Perspective Fields

Planar homography estimation refers to the problem of computing a bijective linear mapping of pixels between two images. While this problem has been studied with convolutional neural networks (CNNs), existing methods simply regress the location of the four corners using a dense layer preceded by a fully-connected layer. This vector representation damages the spatial structure of the corners since they have a clear spatial order. Moreover, four points are the minimum required to compute the homography, and so such an approach is susceptible to perturbation. In this paper, we propose a conceptually simple, reliable, and general framework for homography estimation. In contrast to previous works, we formulate this problem as a perspective field (PF), which models the essence of the homography - pixel-to-pixel bijection. The PF is naturally learned by the proposed fully convolutional residual network, PFNet, to keep the spatial order of each pixel. Moreover, since every pixels’ displacement can be obtained from the PF, it enables robust homography estimation by utilizing dense correspondences. Our experiments demonstrate the proposed method outperforms traditional correspondence-based approaches and state-of-the-art CNN approaches in terms of accuracy while also having a smaller network size. In addition, the new parameterization of this task is general and can be implemented by any fully convolutional network (FCN) architecture.