Visually Classifying between 485 SNES games with 95.8% accuracy

June 03, 2019

In the online class Practical Deep Learning for Coders 2019, we learned how to create a world-class image classifer model for the Oxford-IIIT Pet Dataset, which contains 37 different categories of pets.

I wondered if I could use the same techniques to create a model that could examine a video game screenshot and determine which of ~485 Super Nintendo games it belonged to. Would 485 categories be too much? Would transfer learning from ImageNet work on game screenshots which don’t even exist in the real world?

It ended up working surprisingly well, with an accuracy of 95.8%.

This repository contains:

  • All of the scripts I used to collect the training data from youtube
  • The notebooks I used to create the model
  • My trained model
  • A web front-end
  • Dockerfiles for running the web front-end and inference using CPU
  • A systemd service definition for running the web from
  • Ansible role to clone this repository and set up the systemd service.

Written by Matthew Reishus.

© 2020