datafountain sign detection

分割->抠图->检测

比赛精度要求太高，关键还是使用cascade+更好的backbone，前几大佬基本都用了4个cascade头+cvpr19-HRNet

requirements

python=3.7
    numpy
    pandas
    opencv-python
    pytorch >= 1.0.0
    matplotlib
    pycocotools
    tqdm
    cython
    cffi
    opencv-python >= 4.0
    scipy
    msgpack
    easydict
    matplotlib
    pyyaml
    tensorboardX

数据准备

数据目录为~/data/dfsign，下载zip文件和label文件，解压zip到test和train目录

~/data
├── dfsign
│   ├── test
│   ├── train
│   ├── Test_fix.zip
│   ├── Train_fix.zip
│   ├── train_label_fix.csv

代码准备

代码需要放在特定目录$WORKDIR = ~/working/dfsign

链接数据

ln -s ~/data $WORKDIR/mmdetection

mmdetection extensions

cd $WORKDIR/mmdetection
./compile.sh
python setup.py develop

模型权重准备

将deeplab模型放在$WORKDIR/pytorch-deeplab-xception/run/dfsign
将detection模型放在WORKDIR/mmdetection/dfsign/work_dirs

生成训练数据

cd $WORKDIR/tools
# generate segmentation dataset
python convert2voc.py train_test
# generate detection trainset
python generate_train_chip.py

训练

使用deeplab和mmdetection中cascade

测试

cd $WORKDIR/pytorch-deeplab-xception
# run deeplab
./test.sh

cd $WORKDIR/tools
# crop seg results
python generate_mask_chip.py

cd $WORKDIR/mmdetection/dfsign
# run detect model_1 on images from seg results
python detect.py cascade_rcnn_x101_64x4d_fpn.py work_dirs/cascade_rcnn_x101_64x4d_fpn_1x/9954.pth --chip

cd $WORKDIR/tools
# crop detect results
python generate_detect_chip.py

cd $WORKDIR/mmdetection/dfsign
# run detect model_1 on images from detect results
python detect.py cascade_rcnn_x101_64x4d_fpn.py work_dirs/cascade_rcnn_x101_64x4d_fpn_1x/9954.pth
cd $WORKDIR/tools
# predict_1
python dfsign_submit.py predict_1

cd $WORKDIR/mmdetection/dfsign
# 修改cascade_rcnn_x101_64x4d_fpn_1x.py中25行anchor_ratios为[0.5, 1.0, 2.0]
# run detect model_2 on images from detect results
python detect.py cascade_rcnn_x101_64x4d_fpn.py work_dirs/cascade_rcnn_x101_64x4d_fpn_1x/9946.pth
cd $WORKDIR/tools
# predict_2
python dfsign_submit.py predict_2

# ensemble
python dfsign_ensemble
# 生成的predict.csv为最终结果

DCC-lzhy/dfsign-1