Geolocation estimator from Street View Images
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.
Renames Street View images gathered from API and puts them all in one directory.
Finds all Street View locations on appended map (Images directory) and saves their indices in blue_points_indicies.npz.
Ran k-means with different k, graphed distortion. Currently runs k=22 k-means and saves the model to kmeans22.sav.
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.
A numpy array including index (label) to readable cluster name (location).
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'.
Computes validation accuracy using previous data and hyperparameters. Prints top-1, top-3 and top-5 accuracies.
Loads in a test set to check accuracy per class and highest prediction accuracy for each class. CPU only.
Contains the Google Maps images captured using a macro corresponding to the directory name's latitude.
Used to crop all the images in the latitude directories, cropped images are saved in their corresponding latitude directory.
Appends all cropped images to create a large (~700MB) image of the world map.
BSD 3-Clause License