Using data sources from the games static files we have developed a series of generators for common machine learning vision tasks.
A series of video tutorials showing how to generate data with ACDG and add support for new tracks are avaialble on our wesbite.
In generate_data.py the MultiprocessDataGenerator class can be configured to iterate over samples in a folder made by the interfaces recording harness.
An example configration can be seen in configs/monza.yaml.
We provide a docker compose file to run data generation jobs by running: docker compose --project-directory ./docker/ up -d.
To change which configuration is being used you can modify the line in docker/compose.yaml that sets the CONFIG_PATH environment variable.
For each frame, or a sub sampling of frames, it can generate semantic segmentation maps, normal maps and depth maps from the perspective of the car when a sample was recorded.
To add a new type of data inherit from generator/base.py and register the generator in DATA_GENERATORS in generator/generator.py.
To modify how data is generated you can make changes to the respective DataGenerator.
Each track needs to be manually inspected and a file with associated constants created prior to being able to use the generator, see tracks/monza.py and register the TrackData object in tracks/tracks.py.
Introduction of a new map requires each vertex groups semantic label to be decided on, any vertex groups that should be removed and any that require their materials to be modified so they are distinguishable from an important class.
We usually open the mesh in blender to inspect any classes that are ambiguous and progressively refine them based on visualisation of the coded rules.
Similarly car specific camera positioning data needs to be registered in the cars/ folder.
The details for camera offset and pitch can be found in the game files for each car.
Specifically, each car has a car.ini file inside a data.acd archive that you need to unpack with content manager to access.
After parsing game kn5 files to a .obj file you may need to inspect and adjust certain objects that are semantically different but are textured using the same material.
First import <track_name>.obj into blender.
After making any changes to mesh export it using the following settings:
Then open up the exported .obj in a text editor and use Find and replace to change all the occurrences of o to g .
(Note: This is a bug in blender as vertex groups in obj should have a g prefix)
When importing several different .obj files into blender it will automatically generate two seperate materials if the same name is used in two different objects.
In our case these do refere to the same materials and are genuine duplicates.
To combine these you can go to the scripting tab of blender and run this snippet:
import bpy
mats = bpy.data.materials
for mat in mats:
(original, _, ext) = mat.name.rpartition(".")
if ext.isnumeric() and mats.find(original) != -1:
print("%s -> %s" %(mat.name, original))
mat.user_remap(mats[original])
mats.remove(mat)If you use ACDG in your research or quote our baselines please cite us in your work:
@misc{bockman2024aarkopentoolkitautonomous,
title={AARK: An Open Toolkit for Autonomous Racing Research},
author={James Bockman and Matthew Howe and Adrian Orenstein and Feras Dayoub},
year={2024},
eprint={2410.00358},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2410.00358},
}