Datasets:

PhysInOneP02
/

dataset_card_test

@@ -1,168 +1,511 @@
----
-pretty_name: PhysInOne
-language:
-  - en
-license: cc
-size_categories:
-  - 1M<n<10M
-task_categories:
-  - text-to-video
-  - image-to-video
-  - video-to-video
-  - depth-estimation
-  - image-segmentation
-  - object-detection
-tags:
-  - video
-  - 3d
-  - synthetic-data
-  - physical-reasoning
-  - visual-physics
-  - world-model
-  - video-generation
-  - future-frame-prediction
-  - physical-property-estimation
-  - motion-transfer
-  - multiview
-  - simulation
-  - embodied-ai
-  - mechanics
-  - optics
-  - fluid-dynamics
-  - magnetism
-  - unreal-engine
-  - mpm
-  - sph
-arxiv: 2604.09415
-homepage: https://vlar-group.github.io/PhysInOne.html
----
-# PhysInOne: Visual Physics Learning and Reasoning in One Suite
-[Project Page](https://vlar-group.github.io/PhysInOne.html) | [Paper](https://arxiv.org/abs/2604.09415) | [Code](TODO) | [Dataset](TODO)
-## Dataset Summary
-**PhysInOne** is a large-scale synthetic dataset for visual physics learning and reasoning. It contains **153,810 dynamic 3D scenes** and **2 million annotated videos**, systematically covering **71 basic physical phenomena** across four domains of everyday physics: **mechanics**, **optics**, **fluid dynamics**, and **magnetism**.
-Unlike prior visual physics datasets that are often limited in scale, physical diversity, or annotation richness, PhysInOne provides multi-object and multi-physics interactions in complex 3D environments, together with comprehensive annotations including **geometry**, **semantics**, **motion**, **physical properties**, and **text descriptions**.
-PhysInOne is designed to support research on physics-grounded world models, physically plausible video generation, future frame prediction, physical property estimation, motion transfer, and embodied AI.
-## News
-- **2026-04-10**: PhysInOne paper released on arXiv.
-- **TODO**: Dataset released on Hugging Face.
-- **TODO**: Benchmark code and evaluation scripts released.
-## Dataset Highlights
-| Feature                 | Description                                                                                             |
-| ----------------------- | ------------------------------------------------------------------------------------------------------- |
-| Dynamic 3D scenes       | 153,810                                                                                                 |
-| Dynamic videos          | 2 million                                                                                               |
-| Physical phenomena      | 71                                                                                                      |
-| Physical domains        | Mechanics, Optics, Fluid Dynamics, Magnetism                                                            |
-| Multiphysics activities | 3,284                                                                                                   |
-| 3D objects              | 2,231                                                                                                   |
-| Materials               | 623                                                                                                     |
-| 3D backgrounds          | 528                                                                                                     |
-| Cameras per scene       | 12 fixed cameras + 1 moving monocular camera                                                            |
-| Video resolution        | 1120 × 1120                                                                                             |
-| Frame rate              | 30 FPS; 60 FPS for laser-related scenes                                                                 |
-| Average duration        | Approximately 5.2 seconds                                                                               |
-| Annotation types        | RGB, depth, masks, 3D trajectories, object meshes, material properties, camera poses, text descriptions |
-## Dataset Description
-PhysInOne aims to address the scarcity of large-scale, physically grounded visual data for training and evaluating AI systems. The dataset focuses on visually observable everyday physics and covers four major domains:
-- **Mechanics**: gravity, collision, acceleration, equilibrium, rotation, Hooke's law, conservation of momentum, and related phenomena.
-- **Optics**: reflection, refraction, mirror interactions, laser-related activities, and other visible optical phenomena.
-- **Fluid Dynamics**: liquid motion, splashing, buoyancy, droplets, and related interactions.
-- **Magnetism**: magnetic attraction, magnetic imbalance, and multi-object magnetic interactions.
-Each scene may involve one or more physical phenomena occurring simultaneously or sequentially. These phenomena are instantiated as concrete 3D scenes with diverse objects, materials, and backgrounds.
-## Supported Tasks
-PhysInOne can be used for the following research tasks.
-### 1. Physics-aware Video Generation
-Given text prompts, initial frames, or image conditions, models are expected to generate videos that are not only visually realistic but also physically plausible.
-Potential settings include:
-- Text-to-video generation
-- Image-to-video generation
-- Text-image-to-video generation
-- Video model fine-tuning with physics-rich data
-Suggested evaluation metrics:
-- PMF: Physical Motion Fidelity
-- FVD
-- Human preference / human physical plausibility rating
-### 2. Future Frame Prediction
-Given observed frames from a dynamic scene, models are expected to predict future frames while preserving physically plausible dynamics.
-PhysInOne supports both:
-- **Long-term future frame prediction**: predicting the second half of a video from the first half.
-- **Continuous short-term future frame prediction**: repeatedly predicting the next few frames from streaming observations.
-The dataset supports evaluation from both seen and novel viewpoints.
-### 3. Physical Property Estimation
-Given visual observations of dynamic scenes, models are expected to infer physical properties such as:
-- Young's modulus
-- Poisson's ratio
-- Viscosity
-- Bulk modulus
-- Yield stress
-- Friction angle
-- Initial velocity
-This task is useful for inverse physics, system identification, resimulation, editable scene dynamics, and robot interaction.
-### 4. Motion Transfer
-Given a source video and a target image or scene, models are expected to transfer physically meaningful motion patterns from the source to the target while preserving the target appearance.
-This task is especially challenging because PhysInOne contains complex multi-object and multi-physics interactions.
-## Dataset Structure
-The exact released structure may vary depending on the hosted version. A recommended structure is shown below.
-```text
-PhysInOne/
-├── README.md
-├── metadata/
-│   ├── train.jsonl
-│   ├── val.jsonl
-│   ├── test.jsonl
-│   ├── phenomena_taxonomy.json
-│   ├── material_properties.json
-│   └── benchmark_subsets.json
-├── videos/
-│   ├── train/
-│   ├── val/
-│   └── test/
-├── annotations/
-│   ├── depth/
-│   ├── masks/
-│   ├── trajectories/
-│   ├── camera_poses/
-│   ├── meshes/
-│   └── material_properties/
-└── scripts/
-    ├── download.py
-    ├── load_sample.py
-    └── visualize_sample.py
-```

+---
+pretty_name: PhysInOne
+language:
+  - en
+license: cc
+size_categories:
+  - 1M<n<10M
+task_categories:
+  - text-to-video
+  - image-to-video
+  - video-to-video
+  - depth-estimation
+  - image-segmentation
+  - object-detection
+tags:
+  - video
+  - 3d
+  - synthetic-data
+  - physical-reasoning
+  - visual-physics
+  - world-model
+  - video-generation
+  - future-frame-prediction
+  - physical-property-estimation
+  - motion-transfer
+  - multiview
+  - simulation
+  - embodied-ai
+  - mechanics
+  - optics
+  - fluid-dynamics
+  - magnetism
+  - unreal-engine
+  - mpm
+  - sph
+arxiv: 2604.09415
+homepage: https://vlar-group.github.io/PhysInOne.html
+---
+# PhysInOne: Visual Physics Learning and Reasoning in One Suite
+[Project Page](https://vlar-group.github.io/PhysInOne.html) | [Paper](https://arxiv.org/abs/2604.09415) | [Code](TODO) | [Dataset](TODO)
+## Dataset Summary
+**PhysInOne** is a large-scale synthetic dataset for visual physics learning and reasoning. It contains **153,810 dynamic 3D scenes** and **2 million annotated videos**, systematically covering **71 basic physical phenomena** across four domains of everyday physics: **mechanics**, **optics**, **fluid dynamics**, and **magnetism**.
+Unlike prior visual physics datasets that are often limited in scale, physical diversity, or annotation richness, PhysInOne provides multi-object and multi-physics interactions in complex 3D environments, together with comprehensive annotations including **geometry**, **semantics**, **motion**, **physical properties**, and **text descriptions**.
+PhysInOne is designed to support research on physics-grounded world models, physically plausible video generation, future frame prediction, physical property estimation, motion transfer, and embodied AI.
+## News
+- **2026-04-10**: PhysInOne paper released on arXiv.
+- **TODO**: Dataset released on Hugging Face.
+- **TODO**: Benchmark code and evaluation scripts released.
+## Visual Overview
+<p align="center">
+  <img src="./assets/teaser.jpg" width="900">
+</p>
+<p align="center">
+  <em>PhysInOne covers diverse dynamic 3D scenes across mechanics, optics, fluid dynamics, and magnetism.</em>
+</p>
+## Dataset Examples
+The following examples show representative dynamic scenes from the four major physical domains covered by PhysInOne.
+| Mechanics                            | Fluid Dynamics                   |
+| ------------------------------------ | -------------------------------- |
+| ![](./assets/examples_mechanics.gif) | ![](./assets/examples_fluid.gif) |
+| Optics                            | Magnetism                            |
+| --------------------------------- | ------------------------------------ |
+| ![](./assets/examples_optics.gif) | ![](./assets/examples_magnetism.gif) |
+High-quality MP4 examples are also available:
+| Domain         | MP4                                           |
+| -------------- | --------------------------------------------- |
+| Mechanics      | [View video](./assets/examples_mechanics.mp4) |
+| Fluid Dynamics | [View video](./assets/examples_fluid.mp4)     |
+| Optics         | [View video](./assets/examples_optics.mp4)    |
+| Magnetism      | [View video](./assets/examples_magnetism.mp4) |
+## Dataset Highlights
+| Feature                 | Description                                                                                             |
+| ----------------------- | ------------------------------------------------------------------------------------------------------- |
+| Dynamic 3D scenes       | 153,810                                                                                                 |
+| Dynamic videos          | 2 million                                                                                               |
+| Physical phenomena      | 71                                                                                                      |
+| Physical domains        | Mechanics, Optics, Fluid Dynamics, Magnetism                                                            |
+| Multiphysics activities | 3,284                                                                                                   |
+| 3D objects              | 2,231                                                                                                   |
+| Materials               | 623                                                                                                     |
+| 3D backgrounds          | 528                                                                                                     |
+| Cameras per scene       | 12 fixed cameras + 1 moving monocular camera                                                            |
+| Video resolution        | 1120 × 1120                                                                                             |
+| Frame rate              | 30 FPS; 60 FPS for laser-related scenes                                                                 |
+| Average duration        | Approximately 5.2 seconds                                                                               |
+| Annotation types        | RGB, depth, masks, 3D trajectories, object meshes, material properties, camera poses, text descriptions |
+## Dataset Description
+PhysInOne aims to address the scarcity of large-scale, physically grounded visual data for training and evaluating AI systems. The dataset focuses on visually observable everyday physics and covers four major domains:
+- **Mechanics**: gravity, collision, acceleration, equilibrium, rotation, Hooke's law, conservation of momentum, and related phenomena.
+- **Optics**: reflection, refraction, mirror interactions, laser-related activities, and other visible optical phenomena.
+- **Fluid Dynamics**: liquid motion, splashing, buoyancy, droplets, and related interactions.
+- **Magnetism**: magnetic attraction, magnetic imbalance, and multi-object magnetic interactions.
+Each scene may involve one or more physical phenomena occurring simultaneously or sequentially. These phenomena are instantiated as concrete 3D scenes with diverse objects, materials, and backgrounds.
+## Dataset Structure
+The exact released structure may vary depending on the hosted version. A recommended structure is shown below.
+```text
+PhysInOne/
+├── README.md
+├── assets/
+│   ├── teaser.jpg
+│   ├── dataset_statistics.jpg
+│   ├── annotation_overview.jpg
+│   ├── examples_mechanics.gif
+│   ├── examples_fluid.gif
+│   ├── examples_optics.gif
+│   ├── examples_magnetism.gif
+│   ├── examples_mechanics.mp4
+│   ├── examples_fluid.mp4
+│   ├── examples_optics.mp4
+│   └── examples_magnetism.mp4
+├── metadata/
+│   ├── train.jsonl
+│   ├── val.jsonl
+│   ├── test.jsonl
+│   ├── phenomena_taxonomy.json
+│   ├── material_properties.json
+│   └── benchmark_subsets.json
+├── videos/
+│   ├── train/
+│   ├── val/
+│   └── test/
+├── annotations/
+│   ├── depth/
+│   ├── masks/
+│   ├── trajectories/
+│   ├── camera_poses/
+│   ├── meshes/
+│   └── material_properties/
+└── scripts/
+    ├── download.py
+    ├── load_sample.py
+    └── visualize_sample.py
+```
+Each scene is associated with multiple rendered videos and annotations. A typical metadata entry may look like:
+```json
+{
+  "scene_id": "scene_000000",
+  "activity_id": "activity_0000",
+  "split": "train",
+  "physical_domains": ["mechanics"],
+  "physical_phenomena": ["gravity", "collision"],
+  "caption": "A red ball rolls down a sloped surface and collides with a box...",
+  "duration": 5.2,
+  "fps": 30,
+  "resolution": [1120, 1120],
+  "videos": {
+    "fixed_camera_00": "videos/train/scene_000000/camera_00.mp4",
+    "fixed_camera_01": "videos/train/scene_000000/camera_01.mp4",
+    "moving_camera": "videos/train/scene_000000/moving_camera.mp4"
+  },
+  "annotations": {
+    "depth": "annotations/depth/train/scene_000000/",
+    "masks": "annotations/masks/train/scene_000000/",
+    "trajectories": "annotations/trajectories/train/scene_000000.json",
+    "camera_poses": "annotations/camera_poses/train/scene_000000.json",
+    "meshes": "annotations/meshes/train/scene_000000/",
+    "material_properties": "annotations/material_properties/train/scene_000000.json"
+  }
+}
+```
+## Data Splits
+PhysInOne is split into training, validation, and held-out test sets with an **8:1:1** ratio. To reduce data leakage, 3D assets are assigned exclusively to one split.
+| Split         | Ratio | Description                                                                 |
+| ------------- | -----:| --------------------------------------------------------------------------- |
+| Train         | 80%   | Used for model training and fine-tuning                                     |
+| Validation    | 10%   | Used for validation, ablation, and model selection                          |
+| Held-out Test | 10%   | Used for final evaluation; 3D assets are disjoint from train and validation |
+## Benchmark Subsets
+The paper uses several smaller subsets for efficient benchmarking.
+| Subset                 | Size                  | Used for                                         |
+| ---------------------- | ---------------------:| ------------------------------------------------ |
+| `test-small`           | 772 text-video pairs  | Physics-aware video generation                   |
+| `test-mini`            | 103 scenes            | Long-term and short-term future frame prediction |
+| `test-tiny`            | 20 scenes             | Physical property estimation                     |
+| Motion transfer subset | 273 validation scenes | Motion transfer evaluation                       |
+## Data Fields
+The dataset may contain the following fields.
+| Field                 | Type         | Description                                                              |
+| --------------------- | ------------ | ------------------------------------------------------------------------ |
+| `scene_id`            | string       | Unique identifier of a dynamic 3D scene                                  |
+| `activity_id`         | string       | Identifier of the physical activity                                      |
+| `split`               | string       | Dataset split: train, validation, or test                                |
+| `physical_domains`    | list[string] | Physics domains involved in the scene                                    |
+| `physical_phenomena`  | list[string] | Basic physical phenomena involved in the scene                           |
+| `caption`             | string       | English description of the scene, visual elements, and physical activity |
+| `videos`              | dict         | Paths to videos rendered from fixed and moving cameras                   |
+| `depth`               | path         | Ground-truth depth images                                                |
+| `masks`               | path         | Per-frame object masks                                                   |
+| `trajectories`        | path/json    | 3D trajectories of dynamic objects                                       |
+| `camera_poses`        | path/json    | Camera intrinsics and extrinsics                                         |
+| `meshes`              | path         | Object mesh files                                                        |
+| `material_properties` | path/json    | Physical and material properties of scene objects                        |
+## Annotation Details
+PhysInOne provides synchronized visual and physical annotations for each dynamic 3D scene.
+<p align="center">
+  <img src="./assets/annotation_overview.jpg" width="900">
+</p>
+| Annotation          | Description                                                                                                      |
+| ------------------- | ---------------------------------------------------------------------------------------------------------------- |
+| RGB videos          | Rendered videos from 12 fixed cameras and 1 moving monocular camera                                              |
+| Depth maps          | Ground-truth depth images synchronized with rendered RGB frames                                                  |
+| Object masks        | Per-frame object masks for semantic or instance-level scene understanding                                        |
+| 3D trajectories     | Object-level motion trajectories over time                                                                       |
+| Camera poses        | Camera parameters for multi-view geometric reasoning                                                             |
+| Object meshes       | 3D mesh assets associated with scene objects                                                                     |
+| Material properties | Physical and material parameters such as density, friction, restitution, and other simulation-related properties |
+| Text descriptions   | Human-written and proofread English paragraphs describing the scene and physical activity                        |
+Users should check the released metadata specification for coordinate systems, units, camera conventions, mask encoding, and material parameter definitions.
+## Supported Tasks and Benchmarks
+PhysInOne is designed to support several visual physics learning and reasoning tasks.
+### 1. Physics-aware Video Generation
+Given text prompts, initial frames, or image conditions, models are expected to generate videos that are not only visually realistic but also physically plausible.
+Potential settings include:
+- Text-to-video generation
+- Image-to-video generation
+- Text-image-to-video generation
+- Video model fine-tuning with physics-rich data
+Representative models evaluated in the paper include:
+- SVD-XT
+- CogVideoX-1.5-5B
+- Wan2.2-5B
+Suggested evaluation metrics include:
+- PMF: Physical Motion Fidelity
+- FVD
+- Human physical plausibility rating
+### 2. Long-term Future Frame Prediction
+Given the first half of a dynamic scene, models are expected to predict the second half of the video while preserving physically plausible motion.
+This setting can be used to evaluate:
+- Scene-specific 4D modeling methods
+- Video prediction models
+- Seen-view prediction
+- Novel-view prediction
+Representative models evaluated in the paper include:
+- TiNeuVox
+- DefGS
+- TRACE
+- FreeGave
+- ExtDM
+- MAGI-1
+Suggested evaluation metrics include:
+- PMF
+- PSNR
+- SSIM
+- LPIPS
+### 3. Continuous Short-term Future Frame Prediction
+Given a stream of observed frames, models are expected to continuously predict the next few frames in a real-time or near-real-time manner.
+This setting is useful for:
+- Future-aware robot planning
+- Embodied AI
+- Dynamic scene understanding
+- Short-horizon physical prediction
+Representative models evaluated in the paper include:
+- DefGS
+- FreeGave
+- ExtDM
+- MAGI-1
+Suggested evaluation metrics include:
+- PMF
+- PSNR
+- SSIM
+- LPIPS
+### 4. Physical Property Estimation
+Given visual observations of dynamic scenes, models are expected to infer physical properties of objects and materials.
+Target properties may include:
+- Young's modulus
+- Poisson's ratio
+- Viscosity
+- Bulk modulus
+- Yield stress
+- Friction angle
+- Initial velocity
+Representative models evaluated in the paper include:
+- PAC-NeRF
+- GIC
+This task is useful for inverse physics, system identification, resimulation, editable scene dynamics, and robot interaction.
+### 5. Motion Transfer
+Given a source video and a target image or scene, models are expected to transfer physically meaningful motion patterns from the source to the target while preserving the target appearance.
+Representative models evaluated in the paper include:
+- MotionPro
+- GoWithTheFlow
+Suggested evaluation metrics include:
+- PMF
+- PSNR
+- SSIM
+- LPIPS
+## How to Use
+### Install Dependencies
+```bash
+pip install datasets huggingface_hub
+```
+### Download Metadata Only
+```bash
+huggingface-cli download TODO/PhysInOne \
+  --include "metadata/*" \
+  --local-dir ./PhysInOne
+```
+### Download a Small Subset
+```bash
+huggingface-cli download TODO/PhysInOne \
+  --include "metadata/*" \
+  --include "videos/test-small/*" \
+  --local-dir ./PhysInOne
+```
+### Load with `datasets`
+```python
+from datasets import load_dataset
+dataset = load_dataset("TODO/PhysInOne", split="train")
+sample = dataset[0]
+print(sample.keys())
+print(sample["caption"])
+print(sample["physical_domains"])
+print(sample["physical_phenomena"])
+```
+### Visualize a Sample
+```python
+import json
+from pathlib import Path
+metadata_path = Path("./PhysInOne/metadata/train.jsonl")
+with open(metadata_path, "r") as f:
+    sample = json.loads(next(f))
+print("Scene ID:", sample["scene_id"])
+print("Caption:", sample["caption"])
+print("Videos:", sample["videos"])
+print("Annotations:", sample["annotations"])
+```
+## Data Creation Pipeline
+PhysInOne is constructed through a multi-stage pipeline.
+### 1. Identifying Physical Phenomena and Laws
+The dataset identifies 71 basic physical phenomena across mechanics, optics, fluid dynamics, and magnetism. These phenomena are governed by fundamental physical laws such as Newton's laws, conservation of momentum, Hooke's law, Bernoulli's principle, and related laws.
+### 2. Collecting 3D Assets
+The dataset uses diverse 3D objects, materials, and backgrounds to instantiate realistic physical scenes. These assets include solid objects, interactable objects, destructible objects, deformable objects, granular objects, liquids, materials, and 3D backgrounds.
+### 3. Creating Multiphysics Multi-object 3D Scenes
+Basic physical phenomena are combined into single-, double-, and triple-physics activities. These conceptual activities are then instantiated as concrete 3D scenes with diverse objects, materials, and backgrounds.
+### 4. Simulating Physical Dynamics
+Multiple simulation tools and engines are used to simulate different types of physical dynamics. Chaos Physics in Unreal Engine is used for many everyday rigid-body physical phenomena. MPM is used for deformable and granular objects, and SPH is used for liquid simulations.
+### 5. Rendering and Annotation
+Each scene is rendered from 12 fixed cameras and one moving monocular camera. During rendering, the dataset generates RGB videos and ground-truth annotations including depth, masks, 3D trajectories, camera poses, meshes, material properties, and text descriptions.
+## Intended Uses
+PhysInOne is intended for research on:
+- Physics-aware video generation
+- World models and physical reasoning
+- Future frame prediction
+- Multi-view dynamic scene understanding
+- Physical property estimation
+- Inverse physics and system identification
+- Motion transfer
+- Embodied AI and robot learning
+- Simulation-aware visual representation learning
+## Out-of-Scope Uses
+PhysInOne is not intended to be used as:
+- A perfect replacement for real-world physical measurements
+- A safety-critical physics simulator
+- A dataset for non-visual physics such as thermodynamics or acoustics
+- A benchmark for domains not covered by the dataset
+- A source of guaranteed real-world physical fidelity without validation
+## Limitations
+PhysInOne is a synthetic dataset. Although the scenes are generated with physics simulation and controlled rendering, users should be aware of several limitations:
+- **Synthetic-to-real gap**: Models trained on PhysInOne may require adaptation before deployment on real-world videos.
+- **Simulator approximation**: Current simulators may not perfectly reproduce all real-world physical effects.
+- **Asset distribution bias**: The dataset distribution depends on the collected 3D objects, materials, and backgrounds.
+- **Physics coverage**: PhysInOne focuses on visually observable everyday physics in mechanics, optics, fluid dynamics, and magnetism. It does not cover all areas of physics.
+- **Rendering bias**: The visual appearance is influenced by the rendering engine, lighting, materials, and asset sources.
+- **Storage cost**: The full dataset contains 2 million videos and may require substantial storage and bandwidth.
+## License
+TODO: Add the final dataset license.
+Please also check the licenses of third-party 3D assets, materials, and backgrounds used in the dataset. If individual assets have separate licenses, they should be documented clearly.
+## Citation
+If you use PhysInOne in your research, please cite:
+```bibtex
+@article{zhou2026physinone,
+  title={PhysInOne: Visual Physics Learning and Reasoning in One Suite},
+  author={Zhou, Siyuan and Wang, Hejun and Cheng, Hu and Li, Jinxi and Wang, Dongsheng and Jiang, Junwei and Jin, Yixiao and Huang, Jiayue and Mao, Shiwei and Liu, Shangjia and Yang, Yafei and Song, Hongkang and Wei, Shenxing and Zhang, Zihui and Wang, Bing and Wang, Zhihua and Zou, Chuhang and Yang, Bo},
+  journal={arXiv preprint arXiv:2604.09415},
+  year={2026}
+}
+```
+## Contact
+For questions about the dataset, please contact:
+- TODO: contact email
+- Project page: https://vlar-group.github.io/PhysInOne.html
+## Acknowledgements
+This dataset was created by the vLAR Group and collaborators. Please refer to the paper for the full author list and acknowledgements.