Satellite Image Difference

Description

With a given two images of the same place from the different dates we need to figure out binary mask with changes

Train

Put data as shown below in tree structure of folders
Put pretrained models from Selim in a /models/pretrained folder. We need those for siamese net.
Change paths to the proper one in the config/ files
Run the training proccess with train.py. The first argument should be path to the config file

python train.py config/config_unet++_resnext50.json
python train.py config/config_siamese_seresnext50.json

After that you'll have:
- saved models in /models/saved folder, logs of training
- logs of training proccess in /logs
- predicted non binary masks in /predicted_masks
In the /notebooks/final_submission.ipynb generated a final submission file via averaging outputs from those two models

Solution description

I splited initial large image into small ones applying after that augmentation
Trained Unet++ with resnext50 backbone using Segmentation models on 1 channel image difference
Trained Siamese net with seresnext50 backbone using models architecture from xview2_solution on RGB channel images

Result

Project Organization

├── LICENSE
├── Makefile           <- Makefile with commands like `make data` or `make train`
├── README.md
├── data
│   ├── Images                 <- Original 4 channel images from first and second dates.
│   ├── Images_composit        <- Composed 8 channel images from original images.
│   ├── mask                   <- Binary masks of images differences.
│   ├── Rucode.xls             <- Table to match images from the same location.
│   └── sample_submission.csv  <- Sample submission file.
│
├── models             <- Trained and serialized models
│   ├── pretrained     <- Pretrained models for siamese models from Selim.
│   └── saved          <- Trained on the competition data models.
|
├── notebooks          <- Jupyter notebooks.
│
├── logs               <- Logs of training process: loss, IoU
|
├── config             <- Configuration files for each type of model
|
├── predicted_masks    <- Predicted probability masks in pickle format 
│
├── requirements.txt   <- The requirements file for reproducing the analysis environment,
│                         `pip install -r requirements.txt`
│
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported
├── train.py           <- Training process with evaluation on the end.
├── eval.py            <- Mask evaluation using the trained model.
├── src                <- Source code for use in this project.
│   ├── __init__.py    <- Makes src a Python module
│   │
│   ├── models.py      <- Siamese models from Selim
│   ├── dataset.py     <- Dataset class for satellite images
│   └── utils.py       <- Small preprocess functions like normalization and decoding
│
└── tox.ini            <- tox file with settings for running tox; see tox.readthedocs.io

Project based on the cookiecutter data science project template. #cookiecutterdatascience

maloyan / Space