Traditional methods for motion estimation estimate the motion field F betweena pair of images as the one that minimizes a predesigned cost function. In thispaper, we propose a direct method and train a Convolutional Neural Network(CNN) that when, at test time, is given a pair of images as input it produces adense motion field F at its output layer. In the absence of large datasets withground truth motion that would allow classical supervised training, we proposeto train the network in an unsupervised manner. The proposed cost function thatis optimized during training, is based on the classical optical flowconstraint. The latter is differentiable with respect to the motion field and,therefore, allows backpropagation of the error to previous layers of thenetwork. Our method is tested on both synthetic and real image sequences andperforms similarly to the state-of-the-art methods.
展开▼
