Universal Visual Encoders for Video-Games
Can we learn game representations from images that extract only the games’ content while ignoring its graphic styling? Spoiler: Yes!
Overview of the paper “Contrastive Learning of Generalized Game Representations” by Chintan Trivedi, Antonios Liapis and Georgios Yannakakis presented at IEEE Conference on Games, 2021.
Video games over the past few decades have evolved from using low-bit graphics during the retro/arcade games’ generation to almost having photorealistic visuals in the modern-day AAA titles. However, despite these huge changes in the visual styling over the years, the underlying content of the game that produces these visuals from the game engine remains the same.
This leads us to the question: is it possible to train a single neural network that looks at these game images and extracts just the content of the game while being able to ignore the large variations in the graphical styling? That’s what we explore in this work.
