hub2 / geoguessr-ai

geoguessr-ai

ML project revolving around guessing location by looking at street view image.

Approach

• Download as many street view images as possible, trim and compress

• Create classes using S2 library Used similar partition technique as this paper: https://research.google.com/pubs/archive/45488.pdf

• Split dataset into classes, try to get balanced datasets

• Download a lot more of street view images for small classes, islands, cities and alike (not everything is on google street view as you can see):

• Example image

• Train ResNet-50 with pretrained weights (changed first and last layer to accomodate input and output) using automated script that deploy code and train on https://runpod.io

• Run, test, repeat training. Add well-known augmentations, custom augmentation (PanoramaShifting), more trimming, lower resolution etc.

• Used GradCamPlusPlus (check_net.py) to see what ResNet is looking at to make its decisions

• My best attempt (~geoguessr Master level) used "voting algorithm" on images from multiple years of street view

Things I want to try

• Add multiple heads to guess country/continent and location