SplashSplat: Reconstructing Splashing Liquids from Real-World Multi-View Videos

What changed

Researchers have introduced SplashSplat, a new method for reconstructing splashing liquids from real-world multi-view videos. This approach is designed to handle the transient and view-dependent nature of splashes, which have been largely overlooked in prior reconstruction research that focused on more stable phenomena like smoke or synthetic liquids. To support this research, a new benchmark dataset has been created, featuring 20 real-world scenes of liquid behavior, from gentle streams to violent splashes. These scenes were captured using seven synchronized, calibrated 4K cameras operating at 60 frames per second, with manually refined masks for both the liquid and container. The SplashSplat method itself is based on fusing per-frame liquid Signed Distance Functions (SDFs) derived from masks to establish geometry. A coarse velocity field is then generated using level-set transport between consecutive SDFs. Finally, Lagrangian carriers are advected along this flow, corrected against new observations, and reseeded to decode local Gaussians for differentiable rendering.

Why it matters for builders

This development provides a new tool and dataset for builders working on realistic fluid simulation and rendering. The ability to reconstruct dynamic splashing liquids from real-world data opens up possibilities for creating more immersive visual effects, improving scientific simulations of fluid dynamics, and developing advanced augmented or virtual reality experiences involving water.

Practical impact

SplashSplat demonstrates superior performance compared to state-of-the-art dynamic Gaussian splatting methods on both the new real-world captures and a synthetic benchmark. The method produces physically plausible motion and requires less training cost. Furthermore, the learned representation supports temporal interpolation and style transfer without the need for re-optimization, offering flexibility in post-processing and creative applications.

Caveats and source limits

The provided source is a research paper abstract and excerpt. While it details the methodology and introduces a benchmark, it does not include specific performance metrics beyond a general statement of outperforming existing methods. Information regarding implementation availability, specific hardware requirements, or detailed comparative benchmark results is not present. The publication date is in the future, suggesting this is a pre-print or a projected release.