Phys-Liquid

Abstract

Estimating the geometric and volumetric properties of transparent deformable liquids is challenging due to optical complexities and dynamic surface deformations induced by container movements. Autonomous robots performing precise liquid manipulation tasks must handle containers in ways that inevitably induce these deformations.

We introduce Phys-Liquid, a physics-informed dataset comprising 97,200 simulation images and corresponding 3D meshes, capturing liquid dynamics across multiple laboratory scenes, lighting conditions, liquid colors, and container rotations. We also propose a four-stage reconstruction pipeline to validate the dataset's utility. Experimental results demonstrate improved accuracy in reconstructing liquid geometry and volume, outperforming existing benchmarks.

Dataset Overview

Phys-Liquid systematically simulates liquid deformation governed by the Navier-Stokes equations, providing paired visual observations with accurate geometric supervision.

Figure 1: Simulation samples showing variations in laboratory scenes, lighting, rotation, and liquid properties.

Reconstruction Pipeline

We validate the dataset with a four-stage pipeline: (1) Liquid Segmentation using SAM2 and YOLO-World, (2) Multi-view Mask Generation via diffusion models, (3) 3D Mesh Reconstruction using Triplane methods, and (4) Real-world Scaling.

Figure 7: Overview of our four-step pipeline for reconstructing and scaling 3D meshes of deformable liquids.

Results

Our method, fine-tuned on Phys-Liquid, produces higher-fidelity geometry compared to baselines like InstantMesh and TripoSR, capturing precise morphological features of liquid deformation.

Figure 8: Qualitative comparison of reconstructed meshes by baseline methods and our pipeline.

Citation

If you find this dataset or code useful in your research, please cite our paper:

@inproceedings{ma2026physliquid,
  title={Phys-Liquid: A Physics-Informed Dataset for Estimating 3D Geometry and Volume of Transparent Deformable Liquids},
  author={Ma, Ke and Fang, Yizhou and Weibel, Jean-Baptiste and Tan, Shuai and Wang, Xinggang and Xiao, Yang and Fang, Yi and Xia, Tian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}