MaxFlow: a Convolutional Neural Network Based Optical Flow Algorithm for Large Displacement Estimation

Optical flow estimation is a basic problem in computer vision. FlowNet is the first convolutional neural network based optical algorithm that estimates optical flow by learning the relationship between image pair and the corresponding optical flow. In this paper, MaxFlow is proposed to improve the accuracy of FlowNet. The architecture of MaxFlow is similar to that of FlowNetSimple. MaxFlow uses two kinds of new layers, which are designed specially for estimating large displacements of small scale objects. The new down sampling layer makes the network to predict the maximum displacement in a region. Thus the large movements will not be missed. The new up sampling layer up samples the estimated optical flow fields without using any parameter. It simplifies the network without decreasing the accuracy of the network. Experiments on synthetic datasets and real datasets illustrate that the two new layers are effective and the accuracy of MaxFlow is higher than that of FlowNet.