Guideline

1. Setup environment

Install Anaconda or miniconda to your device. See more information at this link.

# choose your own virtual environment name
conda create --name paddlex_env python=3.8
'''
# if you encountered package errors please consider use
conda create --name paddlex_env python=3.7
'''

conda activate paddlex_env

Config PaddlePaddle environment. Visit this link for more instructions.

# if your device has dedicated GPU
# check the CUDA version on your device
nvidia-smi

# select the suitable CUDA version
# be aware CPU version is also available
conda install paddlepaddle-gpu==2.4.2 cudatoolkit=11.2 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/Paddle/ -c conda-forge

Config PaddleX. The following steps are based on API development mode on UNIX system, for different platform please see more instructions.

pip install paddlex==2.1.0 -i https://mirror.baidu.com/pypi/simple

# for Linux or MacOS
pip install cython
pip install pycocotools

# for Windows, Microsoft Visual C++ 14.0 or greater is required.
# Get it with "Microsoft C++ Build Tools": 
# https://visualstudio.microsoft.com/visual-cpp-build-tools/
pip install cython
pip install git+https://gitee.com/jiangjiajun/philferriere-cocoapi.git#subdirectory=PythonAPI

2. Prepare training dataset

Annotating images used to be iterative and time consuming. This study used EISeg (Efficient Interactive Segmentation), open sourced by PaddlePaddle (see Materials and methods 2.1).

Install the EISeg tool:

pip install eiseg

For instructions on EISeg interactive annotation process, see this link.

A demo of annotated image is available in the directory ./demo dataset. However, the dataset needs to be converted to PaddleX compatible format. For example:

Dataset/ # Dataset directory
|--JPEGImages/ # Raw images
|  |--1.jpg
|  |--2.jpg
|  |--...
|  |--...
|
|--Annotations/ # Annotated images
|  |--1.png
|  |--2.png
|  |--...
|  |--...

|--labels.txt
background
mussel


|--train_list.txt
JPEGImages/1.jpg Annotations/1.png
JPEGImages/2.jpg Annotations/2.png
... ...


|--val_list.txt
JPEGImages/1.jpg Annotations/1.png
JPEGImages/2.jpg Annotations/2.png
... ...

Convert the annotated dataset to PaddleX compatible format. The function is available at the develop branch of the original repository, please see this link.

cd utils
python eiseg2paddlex.py

If you have separated image annotation jobs or want to extend the existing labelling data

cd utils
python merge_coco.py

3. Image augmentation

This step is included in the experiment models (see Python code in the directory ./experiment models).

For detailed information on model API, please see this link.

4. Model training and evaluation

Each model has two sets of scripts where the Bash script (.sh file) configs the parallel GPUs environment; Python script (.py file) includes the:

• Image augmentation
• Dataset path
• Model configurations
• Training hyper-parameters

More details about defining the model segmenters are available at this link.

If your training environment cannot access the Internet for system integrity reasons, please download the pre-trained weights beforehand and direct the path to pre-trained weights in the Python script. For more information please see this link.

# download the Cityscapes pre-trained weights
cd utils
python get_pretain_weights.py

Reach me by email for more discussion. Happy to chat ;)

If you find this work useful for your research, please cite as:

Y.-F. Gu, J. Hu and G.A. Williams, A comparative study
on CNN-based semantic segmentation of intertidal mussel beds, Ecological Informatics
(2023), https://doi.org/10.1016/j.ecoinf.2023.102116

or

@article{Gu_2023,
    doi = {10.1016/j.ecoinf.2023.102116},
    url = {https://doi.org/10.1016%2Fj.ecoinf.2023.102116},
    year = {2023},
    month = {jul},
    publisher = {Elsevier {BV}},
    volume = {75},
    pages = {102116},
    author = {Yi-Fei Gu and Jiaxin Hu and Gray A. Williams},
    title = {A comparative study on {CNN}-based semantic segmentation of intertidal mussel beds},
    journal = {Ecological Informatics}
}