Data generation for the Intuitive Physics Challenge - https://intphys.cognitive-ml.fr
Developed with Unreal Engine 4.18 and the UnrealEnginePython plugin.
The 2.0 version of intphys is a reimplementation of intphys-1.0 based on UE-4.19 using the Python scripting language (intphys-1.0 was developed with UE-4.8 and Lua).
Important note
This software is used for generation of the datasets used in the Intuitive Physics Challenge, it allows users to generate more train scenes but the code for test scenes generation is kept private.
Train samples are always physically possible and have high variability
Test samples have a constrained variability and come as quadruplets: 2 possibles cases and 2 impossibles ones
Each video comes with it's associated depth field and object masking (each object have a unique id), along with a detailed status in JSON format.
This installation process succeeded on Debian stable (Jessie), Ubuntu 16.04 and Ubuntu 18.04. It may be fine for other Unices as well, but this have not been tested.
-
Setup an Epic Games account at https://github.com/EpicGames/Signup/, needed to clone the Unreal Engine repository from github.
-
Setup the compiler. You need
clang-3.8
on Ubuntu 16.04 or below, orclang-3.9
on Ubuntu 18.04 (because 3.8 is no more available on that distribution). The compilation will fail if using a more recent version ofclang
.Make sure you have
clang
installed and that/usr/bin/clang++
effectively points to the 3.8 or 3.9 version. To check it, have a/usr/bin/clang++ --version
, it should output something like:clang version 3.9.1-19ubuntu1 (tags/RELEASE_391/rc2) Target: x86_64-pc-linux-gnu Thread model: posix InstalledDir: /usr/bin
If this is not the case, install it and update
clang
to point to the right version (here 3.9 as example, adapt it for 3.8):sudo apt install clang-3.9 sudo update-alternatives --install /usr/bin/clang clang /usr/bin/clang-3.9 100 sudo update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-3.9 100
-
Install Unreal Engine as documented here.
-
First clone the Engine sources, version 4.19:
git clone https://github.com/EpicGames/UnrealEngine -b 4.19 cd UnrealEngine
-
If compiling with
clang-3.9
you may need to modify a file in the engine sources (the next step may fail, see here for details). Edit the file Engine/Source/ThirdParty/FBX/2018.1.1/include/fbxsdk/core/fbxproperty.h and replace the line 1188 as:return StaticInit(pObject, pName, FbxGetDataTypeFromEnum(eFbxEnum), pValue, pForceSet, pFlags); // return StaticInit(pObject, pName, FbxGetDataTypeFromEnum(FbxTypeOf(*((FbxReference*)0))), pValue, pForceSet, pFlags);
-
Then compile the engine:
./Setup.sh ./GenerateProjectFiles.sh make
-
-
Install dependencies required for intphys dataset generation:
sudo apt-get install libpng++-dev
-
Install intphys in a separate directory. Clone the repository and its UnrealEnginePython submodule from github:
git clone [email protected]:bootphon/intphys.git cd intphys git submodule update --init Plugins/UnrealEnginePython
-
You need Python>=3.6 installed as
/usr/bin/python3.6
(else you need to tweak the UnrealEnginePython plugin here so that it can find python>=3.6 in a non-standard path, for exemple in a virtual environment). You also need the dataclasses module (in the standard library with python>=3.7). To install it on ubuntu 16.04, have a:sudo add-apt-repository ppa:jonathonf/python-3.6 sudo apt-get update sudo apt-get install python3.6 python3.6-dev /usr/bin/python3.6 -m pip install dataclasses
-
Setup your
~/.bashrc
: The intphys code reads the path to UnrealEngine and the path to the Python scripts from theUE_ROOT
andPYTHONPATH
environment variables repectively. Add that to your~/.bashrc
:export UE_ROOT=/absolute/path/to/UnrealEngine export PYTHONPATH=$PYTHONPATH:/absolute/path/to/intphys/Content/Scripts
-
Build
intphys
package: We now need to package the intphys project into a standalone binary program. Just have a:./Tools/build/build_package.sh
-
Use the
intphys.py
program to run the game and generate data. To discover the arguments, have a:./intphys.py --help
-
The basic usage is calling
intphys.py scenes.json -o ./output_data
. This reads the scenes to be generated from thescenes.json
file and write them in the folder./output_data
. -
Generate scenes using the packaged game:
./intphys.py Examples/train.json
-
When developing you may want to run the project within UE4Editor (
--editor
option) or as a standalone game (--standalone-game
option). The--verbose
option is usefull for dev as well. -
Use the
--headless
option to disable direct rendering on screen (the game will be rendered in a frame buffer instead). -
You can also use
Tools/parallel/intphys_parallel.sh
to call several instances ofintphys.py
in parallel and speedup the dataset generation.
In the Tools
directory are stored few utility scripts, see README there for
more information.
Here are some tweaks and error reports usefull for intphys development.
Occured after an update of the nvidia graphic drivers. Just reboot and
try again. sudo glxinfo
helps.
After a system update you may issue an error when running
./Tools/build/build_package.sh
about .gch
precompiled headers. The issue is
like:
fatal error: file '/usr/include/linux/version.h' has been modified since the precompiled header '.../UE4Editor/Development/Engine/SharedPCH.Engine.h.gch' was built
note: please rebuild precompiled header '.../UE4Editor/Development/Engine/SharedPCH.Engine.h.gch'
In that case remove all the .gch
files and recompile the engine:
cd $UE_ROOT
find Engine/Intermediate -type f -name "*.gch" -delete
./GenerateProjectFiles.sh && make