[论文] SplashSplat: Reconstructing Splashing Liquids from Real-World Multi-Vi...

论文概要

研究领域: CV 作者: Peiyu Liu, Dingxi Zhang, Federico Tombari, Marc Pollefeys, Christina Tsalicoglou, Daniel Barath 发布时间: 2026-09-17 arXiv: 2609.20818

中文摘要

飞溅的液体只存在于一瞬间：液膜撕裂成液丝和液滴，外观视角依赖且几乎无纹理，几乎没有能持续足够长时间被追踪的特征。因此，重建研究主要集中在烟雾、合成液体或缓慢变形的表面上。据我们所知，目前不存在溅射液体的同步多视角数据集。因此我们引入了一个包含20个真实场景的基准——从连贯的水流到剧烈的飞溅——由7台同步校准的4K相机以60fps拍摄，附带人工精修的逐视角液体和容器掩码以及固定的评估划分。我们进一步提出 SplashSplat，建立在一个单一原则之上：仅在观察能够约束的地方施加物理结构。从掩码融合得到的逐帧液体 SDF 提供几何信息，相邻 SDF 之间的水平集输运产生粗略速度场，沿此流动平流的拉格朗日载体——根据每个新观察进行校正并在覆盖丢失处重新播种——解码出局部高斯用于可微渲染。SplashSplat 在我们的真实采集数据和合成基准上均优于最先进的动态高斯泼溅方法，运动物理上更合理且训练成本更低。同一表示还支持时间插值和风格迁移，无需重新优化。

原文摘要

A splash lives for a fraction of a second: sheets tear into ligaments and droplets, appearance is view-dependent and nearly textureless, and little persists long enough to track. Reconstruction research has consequently focused on smoke, synthetic liquids, or gently deforming surfaces. To our knowledge, no synchronized multi-view dataset of splashing liquids exists. We therefore introduce a benchmark of 20 real scenes, from coherent streams to violent splashes, captured by seven synchronized, calibrated 4K cameras at 60 fps, with manually refined per-view liquid and container masks and fixed evaluation splits. We further present SplashSplat, built on a single principle: impose physical structure only where the observations can constrain it. Per-frame liquid SDFs fused from the masks provid...

#论文 #arXiv #CV #小凯