Skip to content

deepVector/fcn_satellite

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

43 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

WARNING: The end goal of the segmentation is region proposals for 3D points clouds. The network is not tuned for pixelwise accuracy. Instead, the network is tuned to overpredict roads and buildings because the cost for incorrectly labeling a pixel as background is very high for this specific application. This can be seen in action in the examples directory.

Implementation of a 3 class (background, roads, buildings) semantic segmentation for aerial satelite images using a FCN architecture and Caffe.

Significant portions of the following repositories were used:

  1. https://github.com/shelhamer/fcn.berkeleyvision.org (FCN architecture)
  2. https://github.com/mitmul/ssa (data download and pre-processing)

Path setup:

export PYTHONPATH="[path]/fcn_satellite/helper:$PYTHONPATH"

Data preparation:

cd scripts
# download data
./download.sh
# create merged labels
./merged.sh
# create label txt files for each set
cd [path]/fcn_satellite/data/mass_merged/{train/valid/test}/map
ls *.png>{train/valid/test}.txt
# create input txt files for each set
cd [path]/fcn_satellite/data/mass_merged/{train/valid/test}/sat 
ls>{train/valid/test}.txt
# remove filename from end of txt files above

Download VGGconv weights (optional - generate yourself):

cd ilsvrc-nets
# download weights
vim caffemodel_url # follow link and name convention in file

[ig_fcn8] Train weights (optional):

cd ig_fcn8
python H.py # optional - change the H  matrix
python net.py
python solve.py

[ig_fcn8] Download pre-trained weights:

cd ig_fcn8
# download weights
vim caffemodel_url # follow link and name convention in file

Run model on input image:

python infer.py # change image file names

About

Semantic segmentation of satellite images

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 98.1%
  • Shell 1.9%