Self-Supervised Learning of Object Motion Through Adversarial Video Prediction

Can we build models that automatically learn about object motion from raw, unlabeled videos? In this paper, we study the problem of multi-step video prediction, where the goal is to predict a sequence of future frames conditioned on a short context. We focus specifically on two aspects of video prediction: accurately modeling object motion, and producing naturalistic image predictions. Our model is based on a flow-based generator network with a discriminator used to improve prediction quality. The implicit flow in the generator can be examined to determine its accuracy, and the predicted images can be evaluated for image quality. We argue that these two metrics are critical for understanding whether the model has effectively learned object motion, and propose a novel evaluation benchmark based on ground truth object flow. Our network achieves state-of-the-art results in terms of both the realism of the predicted images, as determined by human judges, and the accuracy of the predicted flow. Videos and full results can be viewed on the supplementary website:\url {https://sites. google. com/site/omvideoprediction}.