README.md

Overall presentation

This is an implementation of the training code of DISIR.

Dataset preprocessing

The training dataset should be stored in a folder MyDataset organized as follows:

• a folder named imgs containing the RGB images.
• a folder named gts containing the ground-truths.

This code does not handle additional data sources such as DSMs.

⚠️ Ground-truth files must have the same names than their associated image.

Example for ISPRS Potsdam dataset:

cd <PotsdamDataset>
sudo apt install rename
cd gts; rename 's/_label//' *; cd ../imgs; rename 's/_RGB//' *

The ground-truth maps have to be categorically encoded (i.e. not in a RGB format):

cd DISIR/train
python format_ground_truth -n <dataset_name> -d <PathToMyDataset>/gts

It is only implemented for INRIA and ISPRS Potsdam datasets but can be easily converted to other datasets. Note that format_ground_truth over-writes the current ground truth maps, it might be wise to make a copy of the initial ones before.

Config files

Configuration files are in train/configs and must be passed as --config argument in the CLIs.

Train a model

python -m semantic_segmentation.train -d ~/data/Potsdam -c configs/conf_potsdam.yml

Export the graph of a trained model using jit

In order to use a model in the QGIS plugin, it has to be saved as a torch sript before. To do so:

python -m semantic_segmentation.export_graph -m data/models/<YourModel>.pth -c configs/conf_potsdam.yml -o <OutputDirectory>