See setup.sh for mamba/conda environment.
After setup run mkdir data to create a data directory where outputs will be stored.
Create a flickr account and obtain an api key using this link. Once you have the key, use it in the following script.
cd src python3 get_flickr_metadata.py
In the script it is only done for NYC, you'll have to use your own shapefiles for custom dataset. This will create a directory named metadata
python3 subsample_dedup_bin.py
This will create another directory named metadata_sdb.
python3 get_images.py
This will create another directory named images.
Not all images will be downloaded, use filter_existing_and_rename.py to only filter down metadata of valid images.
python3 filter_existing_and_rename.py
Once we have the final metadata and images, use satellite_sampler.py to obtain center points of satellite images to be sampled.
python3 satellite_sampler.py
This will give two outputs one is the image data satellite_centers.pkl contain center points to be sampled.
The second thing it contains is satellite center to ground images mapping image_data.pkl. If you can get these two pickle piles you can go to the next step of downloading naip images.
NOTE: If you are not interested in Flickr images and instead want to use some other source for coordinates, create pickle files in format similar to satellite_centers.pkl and image_data.pkl for your data and continue the process down below.
Use download_ortho.py to download the raw ortho iamgery from New York's GIS data website.
The ortho imagery are in format not suitable for deep learning applications. So we first convert it into computer vision format.
run the following three scripts:
python3 convert_orthojp22jpg.py
This converts jp2 images to jpeg images. Then to get geocoordinate bounds of these images run get_ortho_bounds.py
python3 get_ortho_bounds.py
After getting bounds for ortho images, use get_cropped_image_coordinates.py to get cropped image bounds for the corresponding ground images.
python3 get_cropped_image_coordinates.py
Finally, run
python3 get_cropped_images.py
to obtain the cropeed images corresponding to ground images.
Optionally run create_train_test_split.py to get pairs of satellite and ground images. This will create a directory split_m2o with csvs containing paired ground-satellite images.