one-stage-object-detection

one stage object detection include yolov1 yolov2 yolov3 object as point reference https://github.com/xingyizhou/CenterNet

the simple train method which not contain DCN

• modify the code in the models init.py

# from .dlav0 import get_pose_net
#from .resnet_dcn import get_pose_net
# from .pose_dla_dcn import get_pose_net
from .msra_resnet import get_pose_net
__all__=(
'get_pose_net', 

the code need to complile the DCNv2，this code is compliled in windows 10,VS2015 python3.6

the file directory is:

detect.py test the original model(train in coco dataset)

• testvoc.py test the result of train using the PASVOC dataset

data preprocess

• download PASVOC dataset in the data dir move the voc_label.py to the data dir

python voc_label.py

could generate the *.txt like
2007_test.txt
2007_train.txt
2007_val.txt
2012_train.txt
2012_val.txt

cat 2007_train.txt 2012_train.txt 2012_val.txt 2007_val.txt>train.txt

cat 2007_test.txt >val.txt 

train

train code is in the experiments

sh pasvoc_384_origin.sh

using the resnet18dcn the mAP is 0.69 using the resnet50dcn the mAP is 0.73