Intro
This was the final year honours project, the main work focused on improving another student’s previous work on generating Minecraft™ maps using a manual approach and tried to automate it using python scrips and Mapcrafter a 2D renderer of Minecraft maps developed by Moritz Hilscher. The pre-processed data would be fed to the GAN generator and the results then compared with Ridham’s work. Following this possible improvements were proposed before then concluding the dissertation.
The undertaking of this project was split essentially in 2 distinct phases one being spent assessing the viability of the project and the second carrying out what was planned in the first. The main structure of the dissertation as well as any relevant research on the topic was conducted during the first semester. Bi-weekly meetings were organized between me and my module supervisor Ruth Falconer to make sure that the work was running smoothly.
Key Points
Crucial to this was reverse-engineering the Mapcrafter renderer to automate the most important aspect when it comes to generating a GAN’s dataset. Labelling. The idea was to repurpuse the Renderer to create Minecraft isometric maps with color coded biomes. The process was very insighful as it allowed me to get a look at the overall architecture of a medium sized C++ program in the form of a fast multi-threaded 2D renderer.
Following this phase, the images needed to be further retouched using Python to remove some dithering and apply color correction. Finally the set was divided between Training, Validation and Testing set.
The training process took almost an entire day using one of the Custom PCs provided by Abertay thanks to Ruth’s approval. Here are some of the screenshots of the 2D Isometric maps generated by the GAN Mapcrafter.
Link to the Dissertation
You can read the full documentation at this address: Disseration
Samples from the Trained Model
