Unsupervised Domain Adaptation for Optical Flow Estimation

In recent years, we have witnessed significant breakthroughs of optical flow estimation with the thriving of deep learning. The performance of the unsupervised method is unsatisfactory due to it is lack of effective supervision. The supervised approaches typically assume that the training and test data are drawn from the same distribution, which is not always held in practice. Such a domain shift problem are common exists in optical flow estimation and makes a significant performance drop. In this work, we address these challenge scenarios and aim to improve the model generalization ability of the cross-domain optical flow estimation model. Thus we propose a novel framework to tackle the domain shift problem in optical flow estimation. To be specific, we first design a domain adaptive autoencoder to transform the source domain and the target domain image into a common intermediate domain. We align the distribution between the source and target domain in the latent space by a discriminator. And the optical flow estimation module adopts the images in the intermediate domain to predict the optical flow. Our model can be trained in an end-to-end manner and can be a plug and play module to the existing optical flow estimation model. We conduct extensive experiments on the domain adaptation scenarios including Virtual KITTI to KITTI and FlyingThing3D to MPI-Sintel, the experimental results show the effectiveness of our proposed method.