Awesome

SpaceNetUnet

This project is aimed to demonstrate how a deep learning model can be engineered and improoved.
It is based on SpaceNet Vegas data.
Baseline model is Unet like , it is inspired by this blog post.

Model engineering is in notebooks dir

Example outputs are included in the examples directory

Baseline

Improoved model

Download SpaceNet Vegas data

It is implied that you have a valid AWS account and installed cli.
If you don't have cli you can download data from aws s3 public bucket.

aws s3api get-object --bucket spacenet-dataset \
    --key AOIs/AOI_2_Vegas/misc/AOI_2_Vegas_Test_public.tar.gz \
    --request-payer requester AOI_2_Vegas_Test_public.tar.gz  
  
mkdir AOI_2_Vegas  
  
tar -C AOI_2_Vegas -xf AOI_2_Vegas_Test_public.tar.gz AOI_2_Vegas  

Download SpaceNet utilities and install requirements

https://github.com/SpaceNetChallenge/utilities/tree/master
  
pip install -r utilities/python/requirements.txt

Process geo files to images

python utilities/python/createDataSpaceNet.py 'AOI_2_Vegas' \
           --srcImageryDirectory RGB-PanSharpen \
           --outputDirectory 'AOI_2_Vegas_processed' \
           --annotationType PASCALVOC2012 \
           --imgSizePix 650 \
           --outputFileType JPEG \
           --convertTo8Bit  
  
mkdir tiles
  
cp AOI_2_Vegas_processed/annotations/.*jpg AOI_2_Vegas_processed/annotations/.*segcls.tif' tiles  

  

At this point you should have 3851 satellite images and 3851 corresponding building footprint masks.