Combining Convolutional Neural Network And Photometric Refinement For Accurate Homography Estimation

Homography estimation refers to the problem of computing a 3 x 3 matrix which transfers image points between two images of a planar scene or two images captured from the same location. While existing algorithms exploiting hand-crafted sparse image features are well-established and efficient, recent methods based on convolutional neural networks (CNNs) achieve promising results especially for low-texture scenes. This work proposes to solve homography estimation using a hybrid framework HomoNetComb which incorporates deep learning method and energy minimization. In particular, a customized light-weight CNN named HomoNetSim is designed to calculate an initial estimation of homography, where the network is trained in an end-to-end fashion using large amount of image pairs generated from a publicly available dataset. Due to the tiny size of the employed network, the computation time of both training and inference for HomoNetSim can be reduced significantly compared with existing CNN-based homography estimation method. The initial estimate is then refined via gradient-decent algorithm by minimizing the masked pixel-level photometric discrepancy between the warped image and the destination image in a parallel fashion. Extensive experiments on the large scale synthetic dataset demonstrate that the proposed HomoNetComb improves robustness of homography estimation significantly compared with traditional methods based on sparse image features, and meanwhile HomoNetComb achieves a mean average corner error (MACE) of 0.58 pixels which outperforms previous state-of-the-art CNN-based method. Moreover, the usefulness and applicability of the proposed method is demonstrated by applying it to solve a real-world image stitching problem.