StreetViewNet

Geolocation estimator from Street View Images

Root

Data Collection

streetview_pics.py

Uses blue_points_indicies.npz, converts (x, y) points to latitude and longitude and uses the API to download images corresponding to a random point in the .npz file into a given directory.

merge_datasets.py

Renames Street View images gathered from API and puts them all in one directory.

blue_points.py

Finds all Street View locations on appended map (Images directory) and saves their indices in blue_points_indicies.npz.

k-means.py

Ran k-means with different k, graphed distortion. Currently runs k=22 k-means and saves the model to kmeans22.sav.

labeler.py

Resizes and labels Street View images using a given k-means model. The files are renamed to include the label number at the end of their name.

countries.npy

A numpy array including index (label) to readable cluster name (location).

Model

ResNet_Model.py

Training and hyperparameter tuning of ResNet model. Takes in 256x256 images with naming convention "[lat] [long] [label].jpg". Saves the model that achieves the highest validation accuracy as 'net.pt'.

test.py

Computes validation accuracy using previous data and hyperparameters. Prints top-1, top-3 and top-5 accuracies.

analysis.py

Loads in a test set to check accuracy per class and highest prediction accuracy for each class. CPU only.

Images Directory

Latitude Directories

Contains the Google Maps images captured using a macro corresponding to the directory name's latitude.

cropper.py

Used to crop all the images in the latitude directories, cropped images are saved in their corresponding latitude directory.

img_appender.py

Appends all cropped images to create a large (~700MB) image of the world map.

License

BSD 3-Clause License