🌱 PhysGaia: A Physics-Aware Dataset of Multi-Body Interactions for Dynamic Novel View Synthesis

PhysGaia is a pioneering physics-aware dataset for dynamic novel view synthesis aiming beyond photorealism toward physical realism!

[ Project Page | arXiv | Github]

⭐️ Key Highlights

💥 Multi-body interaction
💎 Various materials across all modalities: Liquid, Gas, Viscoelastic substance, and Textile
✏️ Physical evaluation: physics parameters & ground-truth 3D trajectories
😀 Research friendly!!
- Providing codes for recent Dynamic Novel View Synthesis (DyNVS) models
- Supporting diverse training setting: both monocular & multiview reconstruction

📂 Dataset Structure

Each folder is corresponding to each scene, containing the following files:

{material_type}_{scene_name}.zip
│
├── render/                              # Generated images (e.g., avatar renderings)
│   ├── train/                           # Images for training
│   └── test/                            # Images for evaluation
│
├── point_cloud.ply                      # COLMAP initialization (PatchMatch & downsampling)
├── camera_info_test.json                # Monocular camera info for test
├── camera_info_train_mono.json          # Monocular camera info for training
├── camera_info_train_multi.json         # Multi-view camera info for training
│
├── {scene_name}.hipnc                   # Houdini source file (simulation or scene setup)
├── particles/                           # Ground-truth trajectories

👩🏻‍💻 Code implementation

Please check each branch for integrated code for recent DyNVS methods.

Shape-of-motion (arXiv 24.07): https://github.com/mjmjeong/PhysGaia/tree/SOM
4DGS (CVPR 24): https://github.com/mjmjeong/PhysGaia/tree/4DGS_hex
STG (CVPR 24): https://github.com/mjmjeong/PhysGaia/tree/spacetime
D-3DGS (CVPR 24): https://github.com/mjmjeong/PhysGaia/tree/deformable

🔥 Ideal for “Next” Research

🧠 Physical reasoning in dynamic scenes
🤝 Multi-body physical interaction modeling
🧪 Material-specific physics solver integration
🧬 Compatibility with existing DyNVS models

💳 Citation

TBD

🤝 Contributing

We welcome contributions to expand the dataset (additional modality for new downstream tasks, , implementation for other models, etc.) Reach out via opening an issue/discussion in the repo.

🛠️ Future Plans

Update fidelity of the generated scenes
Add more easier scenes: providing more accessible starting points
Add guidelines using Houdini source files: ex) How to obtain a flow field?

💳 License

This project is released under the Creative Commons Attribution-NonCommercial 4.0 license. ✅ Free to use, share, and adapt for non-commercial research