sudo wget -qO- https://get.docker.com/ | sh
or
sudo curl -fsSL https://get.docker.com/ | sh
git clone https://github.com/cbnuirl/cbnu_mmsegmentation.git
cd cbnu_mmsegmentation
sudo docker load -i mmseg_cbnu.tar
sudo docker build –t cbnuirl/mmseg_cbnu:1.0 docker/
sudo docker run --gpus all --name {CONTAINER_NAME} --shm-size=8g –it –v \
{WORK_DIR}:/mmsegmentation cbnuirl/mmseg_cbnu:1.0
If you want to change into container bash:
sudo docker start {CONTAINER_NAME}
sudo docker attach {CONTAINER_NAME}
Change into Cityscapes style custom format like:
data/
└(DATA_NAME)
├gtFine
│ ├train/
│ │ └XX_XX_TXWX_XX_XXX_REXX_XXX_labelIds.png
│ │ -> Grayscaled label image
│ └val/
│ └XX_XX_TXWX_XX_XXX_REXX_XXX_labelIds.png
└leftImg8bit
├train/
│ └XX_XX_TXWX_XX_XXX_REXX_XXX.png
│ -> Original image
└val/
└XX_XX_TXWX_XX_XXX_REXX_XXX.png
Grayscaled label image looks like:
Prepare data like:
data/
└(DATA_NAME)
├leftImg8bit
│ ├train/
│ │ └XX_XX_TXWX_XX_XXX_REXX_XXX.png
│ │ -> Original image
│ └val/
│ └XX_XX_TXWX_XX_XXX_REXX_XXX.png
└label/
└XX_XX_TXWX_XX_XXX_REXX_XXX.png
->MORAI label images that have RGB value
MORAI label image looks like:
Change color_map of tools/convert_datasets/morai.py if categories or segmentation colors are different:
…
color_map = [
[255, 90, 241], # vehicle
[66, 7, 158], # bus
[0, 243, 64], # truck
[248, 158, 235], # policeCar
[211, 222, 241], # ambulance
[255, 255, 255], # schoolBus
[112, 32, 48], # otherCar
[255, 255, 255], # motorcycle
[248, 171, 255], # bicycle
[255, 255, 255], # twoWheeler
[255, 0, 233], # pedestrian
[255, 255, 255], # rider
[9, 161, 181], # freespace
[80, 180, 98], # curb
[118, 78, 176], # sidewalk
[255, 255, 255], # crossWalk
[0, 207, 255], # safetyZone
[134, 59, 141], # speedBump
[248, 28, 81], # roadMark
[255, 188, 123], # whiteLane
[255, 100, 0], # yellowLane
[152, 255, 141], # blueLane
[255, 255, 255], # redLane
[106, 163, 145], # stopLane
[241, 90, 41], # trafficSign
[0, 233, 197], # trafficlight
[0, 182, 255], # constructionGuide
[136, 153, 189], # trafficDrum
[255, 255, 255], # rubberCone
[255, 255, 255], # warningTriangle
[247, 234, 110], # fence
[255, 255, 255], # egoVehicle
[210, 229, 168] # background
]
…
[255, 255, 255]is undefined colors of classes.
TrafficDrum was [136, 153, 179] but grayscaled color is same as safetyZone, so temporary modified to [136, 153, 189] for next procedure.
Change mmseg/datasets/custom.py CLASSES, PALETTE:
CLASSES = ['vehicle', 'bus', 'truck', 'policeCar', 'ambulance',
'schoolBus', 'otherCar', 'motorcycle', 'bicycle', 'twoWheeler',
'pedestrian', 'rider', 'freespace', 'curb', 'sidewalk',
'crossWalk', 'safetyZone', 'speedBump', 'roadMark', 'whiteLane',
'yellowLane', 'blueLane', 'redLane', 'stopLane', 'trafficSign',
'trafficLight', 'constructionGuide', 'trafficDrum', 'rubberCone', 'warningTriangle',
'fence', 'egoVehicle', 'background']
PALETTE = [[255, 90, 241], [66, 7, 158], [0, 243, 64], [248, 158, 235], [211, 222, 241],
[144, 64, 0], [112, 32, 48], [176, 0, 96], [248, 171, 255], [176, 0, 128],
[255, 0, 233], [160, 0, 0], [9, 161, 181], [80, 180, 98], [118, 78, 176],
[48, 16, 0], [0, 207, 255], [134, 59, 141], [248, 28, 81], [255, 188, 123],
[255, 100, 0], [152, 255, 141], [32, 48, 0], [106, 163, 145], [241, 90, 41],
[0, 233, 197], [0, 182, 255], [136, 153, 179], [176, 0, 0], [176, 0, 16],
[247, 234, 110], [0, 0, 0], [210, 229, 168]]
Exit docker bash for GUI. Then, run:
pip install tqdm numpy opencv-python # If not installed
python tools/convert_datasets/morai.py --data_path {DATA_PATH}
{DATA_PATH} would be like data/(DATA_NAME). label folder can be deleted.
Create new configuration file for new data. Location is configs/ocrnet/. Save like 'ocrnet_hr48_512x1024_160k_{DATA_NAME}.py' Modify content:
…
data_root = ‘data/{DATA_NAME}/’
…
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=dict(
type=dataset_type,
data_root=data_root,
img_dir='leftImg8bit/train',
ann_dir='gtFine/train',
pipeline=train_pipeline),
val=dict(
type=dataset_type,
data_root=data_root,
img_dir='leftImg8bit/val',
ann_dir='gtFine/val',
pipeline=test_pipeline),
test=dict(
type=dataset_type,
data_root=data_root,
img_dir='leftImg8bit/val',
ann_dir='gtFine/val',
pipeline=test_pipeline))
For mixed training:
…
dataset_A_train=dict(
type=dataset_type,
data_root=data_root,
img_dir='synthetic/leftImg8bit/train',
ann_dir='synthetic/gtFine/train',
pipeline=train_pipeline)
dataset_B_train=dict(
type=dataset_type,
data_root=data_root,
img_dir='real/leftImg8bit/train',
ann_dir='real/gtFine/train',
pipeline=train_pipeline)
dataset_B_test=dict(
type=dataset_type,
data_root=data_root,
img_dir='real/leftImg8bit/val',
ann_dir='real/gtFine/val',
pipeline=test_pipeline)
dataset_B_val=dict(
type=dataset_type,
data_root=data_root,
img_dir='real/leftImg8bit/val',
ann_dir='real/gtFine/val',
pipeline=test_pipeline)
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=[
dataset_A_train,
dataset_B_train
],
val=dataset_B_val,
test=dataset_B_val)
Mixed training data should be distributed like:
data
└(DATA_NAME)/
├real/
│ ├gtFine/
│ │ ├train/
│ │ └val/
│ └leftImg8bit/
│ ├train/
│ └val/
└synthetic/
├gtFine/
│ ├train/
│ └(val/)
└leftImg8bit/
├train/
└(val/)
Validation set that will not be used can be removed.
NOTE: Original image format(.png or .jpg) and label image format(labelImg.png) should be same between real and synthetic data.
If you want to use pre-trained model to continue training or fine-tuning, add configuration file:
_base_ = …
norm_cfg = …
load_from = “(PRETRAINED_MODEL)”
…
PRETRAINED_MODEL will be like "checkpoints/ocrnet…morai_daegu.pth"