games-cnn

Project Cars (source)

• Predicted Genres: Sports, Racing
• True Genres: Sports, Racing, Simulation

Recognizing game genres from screenshots using convolutional neural networks for multi label learning. For more details see this blog post.

The code is provided mostly for illustrative purposes. It was written with the memory limits of my machine in mind and might not be completely portable.

If you want to classify some images yourself, you can use the script predict/classify_image.py. It takes as arguments a lists of paths to jpg images, which you need to supply, and a path to a keras model and a threshold file.

System Requirements

• Python3
• TensorFlow
• Keras

Repository Structure

preprocess contains some web crawling scripts(get_ids.py,get_data.py) to get the training data from the Steam store.

train contains scripts to create the training data(create_training_data.py) and train a model pretrained on ImageNet(train_pretrained_model.py). It also contains an implementation of bp-mll, see this repository.

predict contains scripts to learn optimal thresholds(get_cutoffs.py), evaluate a model on the test data(evaluate.py) and classify new images(classify_image.py).

cutoffs contains the threshold files for the best model I trained, plus the default thresholds (0.5).