Water Segmentation Competition

The repo is for this competition.

Docs:

• 水务分割识别汇报 slides

Usage

Environment (UNet): pytorch, numpy, tqdm.
Environment (DeepLabV3+): pytorch, numpy, tqdm, tensorboard, opencv-python.

Just following the command below.

Linux (cvmart.net competition)

setup see here

UNet

bash /project/train/src_repo/UNet/train.sh
bash /project/train/src_repo/UNet/inference.sh

DeepLabV3+

bash /project/train/src_repo/DeepLab/train.sh
bash /project/train/src_repo/DeepLab/inference.sh

Linux

UNet

# you must change the path in train.sh and inference.sh
# you can refer to *.bat, or just change randomly it as you want
cd $THIS_REPO_DIR_NAME
bash ./UNet/train.sh [dataset-name]
bash ./UNet/inference.sh

DeepLabV3+

# you must change the path in train.sh and inference.sh
# you can refer to *.bat, or just change randomly it as you want
cd $THIS_REPO_DIR_NAME
bash ./DeepLab/train.sh
python ./DeepLab/inference.py

Windows

UNet

cd %THIS_REPO_DIR_NAME%
UNet\train [dataset-name]
UNet\inference

The [dataset-name] should be Kitti or My, default is My.

DeepLabV3+

cd %THIS_REPO_DIR_NAME%
DeepLab\train
DeepLab\inference

Competition Setup and Usage

This url again if you need to refer.

Download

cd /project/train/
rm -rf src_repo
git clone $THIS_REPO_GIT_OR_HTTP
mv $THIS_REPO_DIR_NAME src_repo
...  # test or train the repo as you want

Train

• go to https://www.cvmart.net/dev/10488/modelDevelopment/train
• click 新建训练任务
• set 执行命令 to bash /project/train/src_repo/DeepLab/train.sh
• do not mark any tick in 预加载模型
• click 提交
• wait till your training is done

Test

• copy the interface by using such as
  mkdir -p /project/ev_sdk/src/
cp /project/train/src_repo/DeepLab/ji.py /project/ev_sdk/src/
• specify the model path in file ji.py
• go to https://www.cvmart.net/dev/10488/modelDevelopment/test
• click 发起模型测试
• click 请选择模型列表, choose the file you trained in the step above
  make sure the model path in ji.py is the same as it here
• click 提交
• wait till your testing is done

Architecture

WaterSegmentation
UNet
    ├─ docs/         <-- some documents
    ├─ model/        <-- model and loss classes
    ├─ utils/        <-- dataset reader, visualization, ...
    ├─ train.py      <-- train function
    ├─ inference.py  <-- inference function
    ├─ ji.py         <-- inference interface
    └─ *.sh / *.bat  <-- quick access
DeepLab
    ├─ docs/         <-- some documents
    ├─ model/        <-- model and loss classes
    ├─ utils/        <-- dataset reader, visualization, ...
    ├─ train.py      <-- train function
    ├─ inference.py  <-- inference function
    ├─ ji.py         <-- inference interface
    └─ *.sh / *.bat  <-- quick access

Interface

Refer to this.

def init() -> nn.Module:
def process_image(
    handle: nn.Module, input_image: np.ndarray,
    args: Any, **kwargs
) -> str

Todo

See Issues.