Global Physical Prior Based Fluid Reconstruction for VR/AR

Abstract

Fluid is a common natural phenomenon and often appears in various VR/AR applications. Several works use sparse view images and integrate physical priors to improve reconstruction results. However, existing works only consider physical priors between adjacent frames. In our work, we propose a differentiable fluid simulator combined with a differentiable renderer for fluid reconstruction, which can make full use of global physical priors among long series. Furthermore, we introduce divergence-free Laplacian eigenfunctions as velocity bases to improve efficiency and save memory. We demonstrate our method on both synthetic and real data and show that it can produce better results.