Abstract
We introduce a self-supervised motion-transfer VAE model to disentangle motion and content from video. Unlike previous work regarding content-motion disentanglement in videos, we adopt a chunk-wise modeling approach and take advantage of the motion information contained in spatiotemporal neighborhoods. Our model yields per-chunk representations that can be modeled independently and preserve temporal consistency. Hence, we reconstruct whole videos in a single forward-pass. We extend the ELBO's log-likelihood term and include a Blind Reenactment Loss as inductive bias to leverage motion disentanglement, under the assumption that swapping motion features yields reenactment between two videos. We test our model on recently-proposed disentanglement metrics, and show that it outperforms a variety of methods for video motion-content disentanglement. Experiments on video reenactment show the effectiveness of our disentanglement in the input space where our model outperforms the baselines in reconstruction quality and motion alignment.
Abstract (translated)
URL
https://arxiv.org/abs/2102.00324
