A Manifold Learning based Video Prediction approach for Deep Motion Transfer

We propose a novel manifold learning based end-to-end prediction and video synthesis framework for bandwidth reduction in motion transfer enabled applications such as video conferencing. In our workflow we use keypoint based representations of video frames where image and motion specific information are encoded in a completely unsupervised manner. Prediction of future keypoints is then performed using the manifold of a variational recurrent neural network (VRNN) following which output video frames are synthesized using an optical flow estimator and a conditional image generator in the motion transfer pipeline. The proposed architecture which combines keypoint based representation of video frames with manifold learning based prediction enables significant additional bandwidth savings over motion transfer based video conferencing systems which are implemented solely using keypoint detection. We demonstrate the superiority of our technique using two representative datasets for both video reconstruction and transfer and show that prediction using VRNN has superior performance as compared to a non manifold based technique such as RNN.