The PyTorch and Keras implementations of wUUNet model designed for multiclass fire segmentation For more information read the paper
The model is tested on Ubuntu 18.04 workstation with NVidia RTX2070
You need to install cuda before installing python packages
sudo python3 ./setup.py install
We have collected a custom dataset of 6250 samples. You can extract it via:
python3 ./dataset.py
In order to train the models run:
python3 ./train_wuunet.py
or
python3 ./train_unet.py
the parameters of training are hardcoded and you can change them directly in training scripts
In order to use *.ipynb evaluation notebooks run jupyter server, e.g.:
jupyter notebook
the source code is reusable for wide range of segmentation tasks as well as extendable by introducing new CNN models to solve the multiclass fire-segmentation task.
Training and evaluation procedures are tightly coupled with storing to/getting data from Filesystem since those files are stored into
${project_root}/output
folder you can increase the perfomance of such tasks via linking the output directory manually to the appropriate dir located in SSD e.g.:
sudo ln -s /ssd/output ${project_root}/output
Model | Binary Jaccard | Multiclass Jaccard |
---|---|---|
UNet FS 224 non-int | 87.43 % | 78.26 % |
UNet FS 224 Gauss | 87.96 % | 79.15 % |
UUNet FS 224 Gauss | 89.92 % | 79.91 % |
wUUNet FS 224 Gauss | 91.35 % | 80.23 % |