Progressive Multi-granularity Analysis for Video Prediction

摘要

Video prediction is challenging as real-world motion dynamics are usually multi-modally distributed. Existing stochastic methods commonly formulate random noise input with simple prior distribution, which is insufficient to model highly complex motion dynamics. This work proposes a progressive multiple granularity analysis framework to tackle the above difficulty. Firstly, to achieve coarse alignment, the input sequence is matched to prototype motion dynamics in the training set, based on self-supervised auto-encoder learning via motion/appearance disentanglement. Secondly, motion dynamics is transferred from the matched prototype sequence to input sequence via adaptively learned kernel, and the predicted frames are further refined through a motion-aware prediction model. Extensive qualitative and quantitative experiments on three widely used video prediction datasets demonstrate that: (1) the proposed framework essentially decomposes the hard task into a series of more approachable sub-tasks where a better solution is easier to be sought and (2) our proposed method performs favorably against state-of-the-art prediction methods.