Generating Videos with Scene Dynamics

TLDR

Large amounts of unlabeled video are leveraged to learn a model of scene dynamics that can be applied to both video recognition and generation. The study aims to learn a scene dynamics model for action classification and future prediction. A spatio‑temporal convolutional generative adversarial network is proposed that separates foreground from background. Experiments demonstrate that the model can generate short videos at full frame rate better than baselines, predict plausible futures from static images, and learn useful action‑recognition features with minimal supervision, suggesting scene dynamics are a promising signal for representation learning.

Abstract

We capitalize on large amounts of unlabeled video in order to learn a model of scene dynamics for both video recognition tasks (e.g. action classification) and video generation tasks (e.g. future prediction). We propose a generative adversarial network for video with a spatio-temporal convolutional architecture that untangles the scene's foreground from the background. Experiments suggest this model can generate tiny videos up to a second at full frame rate better than simple baselines, and we show its utility at predicting plausible futures of static images. Moreover, experiments and visualizations show the model internally learns useful features for recognizing actions with minimal supervision, suggesting scene dynamics are a promising signal for representation learning. We believe generative video models can impact many applications in video understanding and simulation.