PINR: A physics-integrated neural representation for dynamic fluid scenes

Recent research based on neural representation has achieved impressive results in dynamic reconstruction. However, reconstructing high-fidelity dynamic fluid scenes from multi-view videos remains challenging due to the complex fluid phenomena such as translucency, scattering, occlusion, and visual ambiguity. In this paper, we present a novel neural representation that simultaneously exploits physics knowledge and differentiable volume rendering techniques to construct a physically plausible neural field. We first propose a simple yet effective state field to separate fluid and non-fluid contents apart, which essentially reduces the amount of computation and appearance ambiguity. To achieve robust optimization and fast convergence, we introduce intermediate velocity in neural representation based on Helmholtz Decomposition to decouple physical quantities (velocity, pressure and density) that are tightly coupled in physical equations. In contrast to previous physics-informed reconstruction methods, we explore a novel solution by introducing a fast and stable fluid solver to update the intermediate velocity to a physically conserved incompressible velocity. This solver is based on a parallelizable multi-scale architecture, which can effectively avoid degenerate solutions. By adopting the proposed physics-aware loss, we achieved better convergence and superior performance in dynamic fluid scenes. Extensive experiments show that the proposed method can achieve state-of-the-art results for reconstructing dynamic fluid scenes, and the performance and efficiency have been significantly improved in different scenarios.