README.md

Data Prep Steps

Getting the data

Steps to get the dataset are available here

You can also use the bash script get_data <data-dir> to get the entire data from aws and place it in data_dir

Test to see if you are able to read the geojson files

cd spacement/utilities/python/ ./read3band.py

The response should be 251994

Test to see if the bounding boxes are being generated correctly.

cd spacement/utilities/python/

overlay.py

You should see the image below. The white boxes represent the contours and the black boxes represent the bounding boxes. The black bounding boxes are used for training.

Generate the Pascal VOC-Compatible Annotation File

This script will read all the images and the corresponding annotations and generate an Pascal VOC-compatible annotation file.

gen-pascalvoc-format.py

Create the Training/Test split

From caffe root, cd data/spacenet

Execute the script, create-train-val.py

At the end of the execution, we produce trainval.txt and test.txt.

Create the test and train LMDB files

From caffe root, cd data/spacenet.

Execute the script, create_data.sh.

One Last Check

From caffe root,

cd data/spacenet

Check that the file labelmap_spacenet.prototxt has just two labels, building and background as shown below.

item {
  name: "none_of_the_above"
  label: 0
  display_name: "background"
}
item {
  name: "building"
  label: 1
  display_name: "building"
}