AAAI-26 (Oral)
Ke Ma, et al.
Phys-Liquid is a physics-informed simulation dataset of transparent deformable liquids in realistic laboratory scenes.
It provides multi-view RGB images, high-quality liquid masks, metrically scaled 3D meshes, and rich physical metadata to support research on:
The dataset is accompanied by a four-stage reconstruction pipeline and benchmark results against strong baselines such as InstantMesh and TriPoSR.
Extended version (arXiv): (coming soon – arXiv link)
Phys-Liquid contains:
The GitHub repository provides:
We propose a four-stage physics-informed reconstruction pipeline:
Transparent-Liquid Segmentation
Segment the transparent liquid region from RGB images using a specialized segmentation model.
Multi-View Mask Diffusion
Use a diffusion-based model to refine and complete liquid masks across multiple views, enforcing multi-view consistency.
3D Reconstruction
Reconstruct a coarse 3D shape from the multi-view masks using a standard reconstruction backbone (e.g., InstantMesh / TriPoSR).
Mesh Scaling to Metric Dimensions
Apply a physics-informed mesh scaling model to recover metric volume and geometry that match the underlying physical simulation.
The repository contains code for training and evaluating key components, as well as configurations to reproduce the main experiments.
Phys-Liquid enables quantitative evaluation of:
We demonstrate that:
If you find Phys-Liquid useful in your research, please cite:
```bibtex @inproceedings{ma2026physliquid, title = {Phys-Liquid: A Physics-Informed Dataset for Estimating 3D Geometry and Volume of Transparent Deformable Liquids}, author = {Ma, Ke and others}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, year = {2026} }