Flow Supervised Neural Radiance Fields for Static-Dynamic Decomposition

Abstract

We present an approach to synthesize novel views from dynamics scenes captured by multi-view videos of cameras mounted on a driving vehicle. We unify existing methods and propose a new training loss to explicitly disentangle the static background from the dynamic foreground objects using scene flow's magnitude, learnt only from proxy 2D optical flow supervision. We obtain high quality static and dynamic contents separately, which allow us to combine them freely for novel view and time syntheses. We establish a dataset consisting of 5 dynamic scenes with varying difficulties on which we conduct experiments, and show that our method is able to handle challenging scenarios in real-world traffics and create high quality novel view and time syntheses.