auto-seg

• 目前单模型最好效果

dataset	Metric	score(%)	Network
Coco-stuff	mIoU	41.05	ResNeSt101+ASPP+OCR
Cityscapes	mIoU	81.56	HRNet48+OCR
Pascal VOC2012	mIoU	84.26	ResNeSt101+ASPP+OCR
LIP	mIoU	57.55	ResNeSt101+ASPP+OCR
Pascal Context	mIoU	57.72	ResNeSt101+ASPP+OCR
MIT Scene Parsing	mIoU	37.94	ResNeSt101+ASPP+OCR
bdd100k domain adaption	mIoU	53.81	HRNet48+OCR
bdd100k drivable	mIoU	85.06	HRNet48+OCR
SIIM-ACR	Dice	85.20	HRNet48+OCR
airbus	F1-score	83.18	ResNeSt101+ASPP+OCR
Carvana	Dice	99.631	ResNeSt101+ASPP+OCR
KolektorSDD	Dice	93.78	HRNet48+OCR
Inria	mIoU	76.99	HRNet48+OCR
APSIS	Dice	98.41	HRNet48+OCR
Magnetic	Dice	82.06	ResNeSt101+ASPP+OCR
CrackForest	mIoU	63.54	HRNet48+OCR

• 训练和测试命令

  • Train

    CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 tools/train.py --cfg experiments/'your config file' 

  • Test

    python tools/test.py --cfg experiments/'your config file' \
                         DATASET.TEST_SET 'your data list path' \
                         TEST.MODEL_FILE 'your model checkpoint file path' \
                         TEST.SCALE_LIST 'test scale list,default [1]' \
                         TEST.FLIP_TEST True

​