GeoguessrLSTM

Geolocating google street view image sequences using CNN/LSTM.

Map creation

• The geolocation task for this project was restricted to the map of mainland USA
• The map of mainlad USA was extracted as a polygon from the US shapefile. The map of USA mainland polygon is stored in the file infoExtraction/usaPoly.pkl. The python script for extracting the map from the shapefile can be found in the notebook infoExtraction/gridCreation.ipynb.
• The map was then split into grids of roughly even size and the list of polygon grids was stored in the file infoExtraction/usaPolyGrid.pkl. The code for creating the grids can be found in the notebook infoExtraction/gridCreation.ipynb

Dataset

• The dataset for this project was stree view images scraped from random locations per grid. Three images were scraped from a single location using the google street view static API.
• Forty locations were scraped from each grid across 243 grids bringing the total dataset size to 9720 locations.
• The python scripts for data scraping can be found in the notebook infoExtraction/DataScraping.ipynb
• Only a sample of 10 location folders can be found at infoExtration/data/dataCombinedSamples.
• The 10 sample location folder names are split into train and test set of 10 and 2 respectively. This sample train test split can be found at infoExtration/data/trainFiles.npy and infoExtration/data/testFiles.npy respectively.
• Note: The data found in this repo is just a sample dataset to mock training and testing features of the model. The entire dataset can be found at: https://drive.google.com/drive/folders/17nt3Bvcsq-ndvY-uRddk3ZFQpc8EdCzL?usp=sharing

CNN/LSTM model

• The class called Geoguessr is used for training, testing, saving and loading the model. This class can be found in the file machineLearning/geoLSTM.py. The file can be run from terminal using the run instructions provided at the begining of the .py file. The training and testing runs will only be run on a sample of data as there the complete dataset is not in this repo
• The Geoguessr class can also be used by running cells in the notebook machineLearning/geoGuessrLSTM_TrainTest.ipynb to train and test a model.
• The final trained models are not in this repo and can be found at: https://drive.google.com/drive/folders/17nt3Bvcsq-ndvY-uRddk3ZFQpc8EdCzL?usp=sharing