1 IceHrNet

IceHrNet is a Texture-Sensitive Semantic Segmentation method for Zero-Shot transfer learning.

This code contains:

1. dataset(train set: blood cell images stylized by river ice; test set: river ice iamges)
2. train.py/evaluate.py/predict.py
3. PaddleSeg framework (include IceHrNet network, model file: "IceHrNet/PaddleSeg/paddleseg/models/icehrnet.py"; backbone file:"IceHrNet/PaddleSeg/paddleseg/models/backbones/icehrnet_backbone.py")

2 Installation

2.1 Environment Requirements

• OS: 64-bit
• Python 3(3.6/3.7/3.8/3.9/3.10)，64-bit version
• pip/pip3(9.0.1+)，64-bit version
• CUDA >= 10.2
• cuDNN >= 7.6
• PaddlePaddle (the version >= 2.4)

2.2 Install PaddlePaddle

Required Conda:

conda create -n env_icehrnet python=3.8

conda activate env_icehrnet

conda install cudatoolkit=11.7.0 -c conda-forge

conda install cudnn=8.4.1.50 -c conda-forge 

Highly recommend you install the GPU version of PaddlePaddle, due to the large overhead of segmentation models, otherwise, it could be out of memory while running the models.

For example, run the following command to install Paddle with Conda, CUDA 11.7.0

conda install paddlepaddle-gpu==2.5.1 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/

In Python interpreter, run the following command to confirm whether PaddlePaddle is installed successfully

>>> import paddle
>>> paddle.utils.run_check()

# If the following prompt appears on the command line, the PaddlePaddle installation is successful.
# PaddlePaddle is installed successfully! Let's start deep learning with PaddlePaddle now.

# Confirm PaddlePaddle version
>>> print(paddle.__version__)

2.3 Get IceHrNet

Clone the IceHrNet repo from Github

git clone https://github.com/PL23K/IceHrNet.git

2.4 Install PaddleSeg

Run the following command, install PaddleSeg from source. If you make modification to PaddleSeg/paddleseg, it will be efficient without reinstallation.

cd IceHrNet/PaddleSeg
pip install -r requirements.txt
pip install -v -e .

2.5 Verify Installation

In the root of PaddleSeg, run the following command. If there are no error in terminal log, you can use PaddleSeg to train, validate, test and export models with config method.

cd IceHrNet/PaddleSeg
sh tests/install/check_predict.sh

3 Usage

3.1 Pretrained

1. "Fixed Camera River Ice" model (trained by stylized Blood Cell dataset, 0.8512 mIoU). Here.
2. "UAV River Ice" model (trained by stylized Blood Cell dataset, 0.8704 mIoU). Here.

3.2 Predict

cd IceHrNet/src/Transfer/FixedCamera_2_BloodCell_Advanced
python predict.py

or

cd IceHrNet/src/Transfer/UAV_2_BloodCell_Advanced
python predict.py

3.3 Train

cd IceHrNet/src/Transfer/FixedCamera_2_BloodCell_Advanced
python train.py

or

cd IceHrNet/src/Transfer/UAV_2_BloodCell_Advanced
python train.py

3.4 Evaluate

cd IceHrNet/src/Transfer/FixedCamera_2_BloodCell_Advanced
python evaluate.py

or

cd IceHrNet/src/Transfer/UAV_2_BloodCell_Advanced
python evaluate.py

4 Citation

If you find our project useful in your research, please consider citing:

@misc{icehrnet2023,
    title={IceHrNet, Texture-Sensitive Semantic Segmentation method for Zero-Shot transfer learning},
    author={Zhiyong Yang},
    howpublished = {\url{https://github.com/PL23K/IceHrNet}},
    year={2023}
}