This code accompanies the paper Densely Annotated Photorealistic Virtual Dataset Generation for Abnormal Event Detection

• GTAVisionNative (compile from source here or download x64 here)
• ScriptHookV (download here)
• ScriptHookVDotNet V3 (download here)
• Menyoo GTA5 Trainer (download here)
• VS2019
• others managed by Nuget

git clone https://github.com/RicoMontulet/GTA5Event.git

• Open GTA5Event.sln in VS2019
• Right click GTA5Event in Solution Explorer: properties -> Build Events and in Post Build event command line add: COPY "$(TargetPath)" "path/to/Grand Theft Auto V/Scripts"
• Right click Solution 'GTA5Event' itself and go to: properties -> Configuration Properties and set the configuration to Release for x64
• Right click solution and click Build Solution

• After you compile the GTA5Event, copy compiled files from the GTA5Event/bin/Release to path/to/Grand Theft Auto V/Scripts.
• Modify the gta_config.ini file, update the directories + move the file to the path/to/Grand Theft Auto V/Scripts.

• Start the game
• Press "PageUp" to activate plugin
• F3 to enter the free camera mode
• F10, F11 and F12 to change states of the plugin.
• Use Z to disable and enable ingame notifications (disable notifications before collecting data)

• In Display
  • set Radar to Off
  • set HUD to Off
  • set game to Windowed mode
• In Graphics
  • set MSAA to Off
  • set Pause Game on Focus Loss to Off
  • set Frame Scaling to off
• In Notifications
  • set all notifications to Off

When this state is active you can scout the game to find locations, using "[ ]" keys you can cycle through already existing locations. When you find a new location press "L" to save it, a popup will ask for a name so you can find it again later. Make sure these names are unique.

When this state is active you can create a ROI for the currently selected location, ie use state F10 to select a location with "[ ]" keys and press F11. If you have a ROI saved already for the current location it will show you this in game. If not, you can use the mouse to position the sphere where you want a corner of the ROI to be. Use "U" to update the ROI (note: ROI is defined as a closed polygon). You can use "K" to remove the last corner point in case you messed up.

When this state is active it will put the camera in the first location and start the data collection process. It will automatically cycle through all available locations and make a new folder "locations_processed" where it will save a new json containing the FPS, cluster centers and locations of people. Make sure the game screen is in front of any other windows since screenshots are used for RGB data. It's best to now alt-tab out of the game (so you don't move the mouse accidentally) and leave the pc alone until it's done gathering data.

PageUp : Enable the plugin and load previously created scenes

• ANY STATE:
  • Z : toggle on and off in game notifications (disable notifications when collecting data)
  • N : Save all locations to disk (folder is specified in gta_config.ini)
• F10:
  • L : Add new location, you will be promted for a name.
  • [ : Go to previous location
  • ] : Go to next location
• F11:
  • U : Update ROI polygon with the sphere's location
  • K : Remove last ROI point from current location

To load the data in python please have a look in the post processing code.

• 0 : Environment objects like floor, stairs, buildings
• 1 : Persons
• 2 : Cars and trucks
• 3 : Waving flags, plants, trees
• 4 : Beach sand, grass
• 7 : sky

Please look at the post processing code to find out how to use the depth information since these are NDC values, not meters!

The location json contains:

• location name
• camera position
• camera rotation
• player position
• player rotation
• ROI of the location
• PedGroups: contains the initial positions of all the people that belong to this group
• PedCenters: contains the original centers around which the people where spawned
• PedIdGroup: contains the handles of all the spawned peds, and the cluster center they belong to
• fps: frames per second the video was rendered at
• Action: the action that happens at frame 300
• Current time: is the time that the video was recorder in game time

The frame json contains:

• Currentweather: the weather of that scene [0=ExtraSunny, 1=Clear, 2=Clouds, 3=Smog, 5=Overcast, 6=Raining, 13=Snowing, 7=Thunderstorm]
• CamFOV: the field of view of the recording camera (always 50)
• Detections: List of all people in the scene with they type (1=person), distance to the camera in meters, their handle which is a unique id, and a list of their bones containing on-sceen relative x, y [0, 1] coordinates

• Borrowing code from umautobots
• Borrowing post processing code from racinmat
• Thanks to H2020 Mindspaces for funding
• Publishing data from game according to Rockstar Games

Please cite us :)

@inproceedings{montulet2021densely,
  title={Densely Annotated Photorealistic Virtual Dataset Generation for Abnormal Event Detection},
  author={Montulet, Rico and Briassouli, Alexia},
  booktitle={Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10--15, 2021, Proceedings, Part IV},
  pages={5--19},
  year={2021},
  organization={Springer International Publishing}
}