s22-team10-project

GEOGUESSR PROJECT

Team Members:
Andrew Preston
Matt Fadler
Riley Welch
Yassaman Nassiri
Jansen Long

For a step-by-step guide to the project please see:

DEMO

All files can be found under the folder geoguessr

Overview

---

Grid Making

We loaded a cartographic boundary shapefile of the U.S. available through the US Census Bureau (Source) into mainlandGrid.ipynb.

Next we isolated the mainland U.S.

We overlayed it with a grid and combined the two to cut off the grids using the boundry of the U.S. A function called MergeFactor was used to combined sections that were too small after the previous process.

This left us with a file called mainlandGrid.pickle that was a map of the mainland U.S. with 65 grids across it. Each section of the grid was associated with a number 0-64.

Data Scraping

Using this map data and datascraper.ipynb we scraped images from Google maps using the Google maps API. For each grid section we scrapped 10 locations (650 location total). For each location we got 3 pictures (1950 pictures total), since Google map images are 360-degree images we used the API to grab an image of size 400px by 200px at 0, 90, and 180-degrees. For Training/Testing we used an 80%/20% split; 8 locations or 24 images per section for training and 2 locations or 6 images per section for testing. The format for labeling each image is [grid no.]_[unique no.].jpg (ex. 0_5.jpg).

Source: Used this as a reference for the grid maker and datascraper.

Prepping Data

After scrapping the data we were left with two folders training_data with 1560 unique images and testing_data with 390 unique images. In-order to be usable by the model we needed to convert the images into numpy arrays. Using the load_data.ipynb file we converted training_data images into nparray x_train and testing_data images into nparray x_test. We also created nparray y_train and y_test that held the grid section number for each image in its corresponding nparray. This tuple of nparrays were inspired and modeled after the MNIST nparrays used in OLA8 from class.

Model Training

Using the x_train and y_train arrays we trained up a Convolution Neural Network (CNN). After sucessfully training, the trained model was saved in a '.h5' file. Saving this model this way allows for the reuse of the trained model without losing the weights of the model.

Output

Using the trained model, the x_test data can be loaded into the model and ran. Using the predict() function on the model we can isolate a grid section to see the weights the model predicts it to be. Using the argmax() function on the array of weights gives the index location on the array with the highest weight. This highest weight represtents the best guess given by the model. This index number is then translated into the corresponding grid section the model is guessing. This predicted grid number can be compared to the grid number entered by the user in the predict() function to see if they match. If they do not match you can see how far away the sections are using the Numbered Grid below. We can also output the pictures from the actual grid and predicted grid to compare them visually.