The config file is a mmcv style configuration file. Please first prepare the nuScenes datasets and specify the data path carefully in config. Then specify the corruptions parameters using the below format, our nuScenes-C dataset use the following severity:
corruptions = [dict(type='CameraCrash', easy=2, mid=4, hard=5),
dict(type='FrameLost', easy=2, mid=4, hard=5),
dict(type='MotionBlur', easy=2, mid=4, hard=5),
dict(type='ColorQuant', easy=1, mid=2, hard=3),
dict(type='Brightness', easy=2, mid=4, hard=5),
dict(type='LowLight', easy=2, mid=3, hard=4),
dict(type='Fog', easy=2, mid=4, hard=5),
dict(type='Snow', easy=1, mid=2, hard=3)
]Support corruption types include MotionBlur, Fog, Snow, ColorQuant, Brightness, LowLight, CameraCrash, FrameLost. By default, all the corruptions include three different levels (e.g., easy, mid, and hard). The number should between 1 and 5.
The corruption generation relies on third-party packages, install by running:
pip install imagecorruptionsAll the corruption types are defined at corruptions.py. You can follow the script to define additional corruption types.
To generate data, run the following command:
cd corruptions
bash tools/generate_dataset.sh <CONFIG>This might take several days, you can generate different corruption types parallelly to accelerate.
