ContPhy-Gen Codebase & Tutorial

Codebase for ContPhy dataset generation.

ContPhy: Continuum Physical Concept Learning and Reasoning from Videos

Zhicheng Zheng*, Xin Yan*, Zhenfang Chen*, Jingzhou Wang, Qin Zhi Eddie Lim, Joshua B. Tenenbaum, and Chuang Gan (* denotes equal contributions)

ICML 2024
Links | Project Page | Paper (Arxiv) | Dataset Download | Cite ContPhy

---

Contents

Generation Tutorial | Postprocessing | Read Data | Dataset Notes | Manual Bug Fixes | Citation

Tutorial: Generating ContPhy Dataset & Scaling-up Customized Dataset

Using Unity3D Editor

If you are a beginner to Unity/C# development, familiarizing with the ContPhy dataset generation and customization logic through the user-friendly Unity Editor could quickly get you started and facilitate advanced programming.

The steps to run generation pipeline:

1. Install Unity Editor.

• Download Unity3D version 2021.3.17. Recommend to install in Unity Hub.

• (Optional) Install DOTNET sdk, install VSCode C#/Unity plugins.

2. Start a new project.

• Open Unity Hub, start a New Project.

• Choose a HDRP (High Definition Render Pipeline) template, for example, a blank HDRP template named High Definition 3D.

• Name the project, for example, TestContPhy.

3. Import dependencies.

Note: The complete ContPhy dataset generation depends on some commercial packages, whose legitimate licenses should be obtained at the Unity Asset Store. Add or remove dependencies according to the requirement of your customized dataset. One can also modify the codebase to fit in open-source alternatives.

Package List

• ContPhy-Gen.

  a. Move all the files in this repository to the directory "TestContPhy/Assets/", like this,

  cd path/to/the/project/TestContPhy
  git clone git@github.com:zzcnewly/ContPhy-Gen.git
  mv ./ContPhy-Gen/* ./Assets
  rm ./ContPhy-Gen

  b. Install ffmpeg and set its path in PlacementConfig.cs.

  c. Edit the path to your bash (Linux/Unix) or cmd (Windows) in RunCommandLine.cs

• Unity Perception 1.0.0-preview.1

  Setup requires several steps.

  • Install Perception Package. In the toolbar, find Package Manager -> click + -> from git url -> enter com.unity.perception to install version 1.0.0-preview.1

  • In the toolbar, find Edit -> Project Settings -> Editor -> Asynchronous Shader Compilation, uncheck it.

  • In the toolbar, find Edit -> Project Settings -> Perception -> Base Path, change to TestContPhy/output/images. If no existing folder, create one. If no Perception option, try with restarting the editor.

  • In Assets/Settings/HDRP High Fidelity.asset in the Project window, set Lit Shader Mode to Both.

• Computational Geometry Unity Library:

  Required. From Github. Move Assets/_Habrador Computational Geometry Library to the Assets folder.

• Unity Mesh Simplifier

  Required. From Github. Go to the Assets folder. Then git clone the repo.

Optional Packages

These packages are optional when generating a new customized dataset but required when generating ContPhy dataset.

Note: Missing the following packages may lead to compile errors. If you do not require a specific package, comment out its associated code by "//".

• Mesh Voxelizer 2.4.1

  The tool is used to convert object meshes to 3D point clouds. Import from Unity Assets Store.

  If you do not require generating 3D point clouds, comment line 13, 99-112 in BaseManager.cs.

  Actually there are some free tools that can serve as an alternative, for example, unity-voxel at Github.

• Filo - The Cable Simulator 1.4

  For the scenario "rope-pulley system". Import from Unity Assets Store.

  Manual Bug Fixes: Check here. There are some bugs in the original codebase, we need to fix them by ourselves.

• Obi Fluid 6.5.1 + Obi Cloth 6.5.1 + Obi Softbody 6.5.1

  3 packages respectively serve for the scenario "fluid", "cloth", and "soft ball". Import from Unity Assets Store, better to follow the sequence above.

• Ignis - Interactive Fire - URPHDRP v2.1.6

  For fire simulation. Import from Unity Assets Store. We developed a fire physical reasoning scenario based on this fire simulation engine.

  Manual Bug Fixes: Version 2.1.6 is not bug free. Check Manual Bug Fixes to fix bugs.

---

4. Check and run generation

• Restart the Unity Editor.

• Check there is no compile error now. If the error is from the scenario or code unrelated to your needs, comment the source code to remove the error reports.

• According to which scenario data is to be generated, modify the class name in MainDataGeneration.cs (line 16).

• In editor Project window, open the file Assets/Scenes/main.unity.

• In editor Game window, choose the image quality, here we choose FullHD(1920x1080).

• Click the triangular button play to run generation. Click the square button stop to stop generation.

• Check dataset in TestContPhy/output/. If having error generating videos, delete TestContPhy/output/images/solo folder and try to run again.

• explore and modify generation parameters in the files under TestContPhy/Assets/Scripts/Configs, for example,

  // generate 2000 videos with complete annotations/sensor-outputs
  public const int ValidIterNum = 2000;
  // set a random seed for generation
  public const int randomSeed =5834;
  // set dataset folder name
  public const string pre_name = "fluid_slides";
  ...

• Close editor Scene window to slightly accelerate rendering speed. For higher performance, you may want to deploy the project to cluster server.

Using Cluster Server

If you are an experienced Unity/C# developer, you can scale up the dataset on servers, e.g. high-performance computing clusters. But note that the simulation and rendering jobs in this codebase have not been optimized for GPU computing and demand more CPU resources than GPU.

The launch command might be like this:

path/to/Unity/Hub/Editor/2021.3.17f1/Editor/Unity  -openfile "path/to/TestContPhy/Assets/Scenes/main.unity" -executeMethod EnterBatchMode.PlayScene

For other tips, check Using Unity3D Editor.

Postprocessing

After generation, run postprocessing script to filter valid trials.

python path/to/TestContPhy/Assets/Scripts/Python/postprocess.py --origin path/to/original/data/folder --output path/to/processed/data/folder

# for example
cd path/to/TestContPhy
python ./Assets/Scripts/Python/postprocess.py --origin ./output/fluid_slides --output ./output/fluid_slides_new

Read and Visualize Annotations

The annotations are provided in a human-readable JSON format. Additionally, a demonstration script is included to facilitate loading the dataset and visualizing its structure and dimensions.

Note: Check data structure and shapes in Print.md.

# for example,
# only print shapes
cd path/to/TestContPhy
python ./Assets/Scripts/Python/read_and_visualize_data.py --trial_path ./output/fluid_slides_new/0 --print_shapes

# print shapes, plot the point clouds for each frame, and directly show a selected frame.
python ./Assets/Scripts/Python/read_and_visualize_data.py --trial_path ./output/fluid_slides_new/0 --print_shapes --visualize_frame --plot_save_path ./output/images/cache --selected_frame 40

Dataset Notes

We are working with .mp4 formatted videos captured at 30 frames per second, comprising 500 videos with fixed lengths specific to each scenario: 250 frames for fluids, 150 for ropes, 145 for cloths, and 120 for balls. However, upon decoding the videos into frames and running the data loading code to fetch the sensor data, please observe the following guidelines:

1. If there is a discrepancy of 2 frames fewer than the specified numbers for any scenario, align the $n^{th}$ frame of each video with the $(n+2)^{th}$ index of the corresponding sensor data.

2. In instances where there is a variation in the number of frames for particle/mesh sensor data among different objects within the same video—particularly noticeable when the duration of cloth frames is not synchronized with other objects—consider the $30^{th}$ frame as the start time for the cloths.

3. We have identified a bug where rigid objects may display incorrect poses in the cloth scenario. We will address and resolve this issue shortly.

Should you have any confusions or problems, please open an issue under this repo. Thanks!

Manual Bug Fixes

# If the package "Filo" is imported, fix the following bugs.
# In the file "TestContPhy/Assets/FiloCables/Scripts/Cable.cs"

# Line 353
- float distance = link.body.SurfaceDistance(t1.Value, t2.Value, !link.orientation, false);
- link.body.AppendSamples(sampledCable, t1.Value, distance, 0, false, link.orientation);
+ //Fixed for consistancy of rope length when rolling
+ link.body.AppendSamples(sampledCable, t1.Value, link.storedCable, 0, false, link.orientation);

# Line 696
+ if (body.name != currentJoint.body1.name && body.name != currentJoint.body2.name)
// Only split if the body is a disc or a convex shape, and the raycast hit is sane.
if ((body is CableDisc || body is CableShape) && hit.distance > 0.1f && hit.distance + 0.1f < currentJoint.length)

# Line 804
// Sample the link, except if the cable is closed and this is the first link.
if (!(i == 0 && closed) || links[i].type == Link.LinkType.Attachment || links[i].type == Link.LinkType.Pinhole)
+ {
+     // Rolling links (only mid-cable)
+     if (links[i].type == Link.LinkType.Rolling)
+     {
+         Vector3? t1 = null, t2 = null;
+         if (prevJoint != null)
+             t1 = prevJoint.body2.WorldToCable(prevJoint.WorldSpaceAttachment2);
+         if (nextJoint != null)
+             t2 = nextJoint.body1.WorldToCable(nextJoint.WorldSpaceAttachment1);
+         if (t1.HasValue && t2.HasValue)
+         {
+             float distance = links[i].body.SurfaceDistance(t1.Value, t2.Value, !links[i].orientation, false);
+             if (links[i].storedCable < distance){
+                 var link0 = links[i];
+                 link0.storedCable = distance;
+                 links[i] = link0;
+             }
+         }
+     }
    SampleLink(prevJoint, links[i], nextJoint);
+ }

# If the package "Ignis" is imported, fix the following bug.
# In the file "TestContPhy/Assets/OAVA-Flame/Scripts/Engine/FireTrigger.cs"
Line 137
- Vector3 closestPoint = other.ClosestPointOnBounds(transform.position);
+ FireRecorder fr = flameObj.GetComponent<FireRecorder>();
+ Vector3 closestPoint = (fr != null&&fr.touchPoints.ContainsKey(other.name))? fr.touchPoints[other.name]: other.ClosestPointOnBounds(transform.position);

Citation

Welcome to cite ContPhy if you find the paper, dataset, and implementations useful in your research :)

@inproceedings{zheng2024contphy,
  title={ContPhy: Continuum Physical Concept Learning and Reasoning from Videos},
  author={Zheng, Zhicheng and Yan, Xin and Chen, Zhenfang and Wang, Jingzhou and Lim, Qin Zhi Eddie and Tenenbaum, Joshua B and Gan, Chuang},
  booktitle={International Conference on Machine Learning},
  year={2024},
  organization={PMLR}
}