Automated Building Detection using Neat-EO.pink

Network & setup

Network details and how to setup can be found here: https://neat-eo.pink. Advised to walk through the tutorial.

Dataset - xView2 Challenge

1. Download raw dataset:

The xView2 dataset can be downloaded from https://xview2.org/dataset (login required).

2. Extract raw dataset:

Extract the contents of the downloaded .tar file:

tar -xvzf <file.tar>

Place the extracted 'images' and 'labels' folders in the 'data/xview' folder. For training and testing the model all data releases are combined.

3. Create Training Datset:

The training dataset can be created by executing the following command:

python neat-eo/preprocess_xview.py --config config.toml --crop 512 512

The above command will create the dataset as per the specifications and splits the 1024 x 1024 images into four 512 x 512 images.

Configuration:

preprocess_xview.py accepts the command line arguments described below,

usage: preprocess_xview.py [-h] --config CONFIG --crop WIDTH HEIGHT 


optional arguments:
  -h, --help            Show this help message and exit (NEEDS TO BE ADDED)
  --config CONFIG       Path to config file
  --crop                Crops image into smaller images of specified width and height 
                        (significantly increases processing time)

Initial results

Model:

Stored on sharepoint

Trained on:

train.taras downloaded from https://xview2.org/dataset

Tested on:

Images corresponding to nepal-flooding disaster which were not part of train.tar.

Initial Results_:

Mean IOU: 0.6927478022575819

Example 1:

Image	Label

Prediction	Diff map

Example 2:

Image	Label

Prediction	Diff map

Example 3:

Image	Label

Prediction	Diff map

Next steps

...