Content-based Game Classification with CNN

This page is for a demo about my recent work on using convolutional neural network for content-based game classification. If you like the demo and interested in knowing more, read my post here.

The graph is the result of classifying about 100 games according to whether they are real-time strategy games or not. I used a dataset of images extracted from gameplay videos for training. Each point in the graph is a 5-minute video clip of a game. Nodes of the same colour belong to the same game and there are 10 nodes per game. Each node is connected to 10 of its nearest neighbours according to the first two components of t-SNE, calculated on the results of a deep convolutional neural network trained to classify real-time strategy games from 2M gameplay images. You can zoom-in to see the game titles and to see what games are similar to each other (with respect to their RTS elements). Drag a node to see its neighbours more clearly. Again, this is the very short description about the demo and you can continue reading here.

If you wait a few minutes, you will see games clusters like this:

And you can see it in fast motion:

This is still a work-in-progress and I will keep adding content and updating the demo as I go.


If you have any question, please send me an email.

Template design by Mohammad Shaker, Strong Emotions ©, 2015