boboyes/leagueoflegends-minimap

League of Legends Synthetic Minimap Dataset (Sample)

⚠️ Dataset Note

This repository contains a small sample (1000 images) of the synthetic dataset.

To generate the full image dataset of any size used for training high-performance models, please use the generator script available in my GitHub repository:

👉 [my_github]

Dataset Description

This dataset consists of synthetically generated League of Legends minimap images designed for training object detection models (specifically YOLO). It utilizes a complex rendering pipeline to simulate various game states, vision conditions, and champion positions.

Key Features

• Fog of War Simulation: Randomly generated fog masks to simulate limited vision.
• Map Objects: Dynamic placement of Towers, Inhibitors, Nexus, Jungle Monsters, and Baron/Dragon.
• Game Effects: Simulates Recall (Blue/Red), Teleport, Ping waves, and other artifacts to mimic noisy real-world gameplay.
• Observer Viewport: Generates a white "camera" rectangle simulating the observer mode.
• Augmentations:
  • JPEG Compression Noise: Enabled by default to mimic stream artifacts.
  • Icon Overlap: Champions can cluster together (simulating teamfights).
  • Background/Icon Augments: Blur, downscaling, and color distortion options.

File Structure

• train/, val/, test/: Image splits.
• labels/: Standard YOLO labels (class_id x_center y_center width height).
• viewport_labels/: Coordinates for the observer camera rectangle (x y w h in pixels).
• data.yaml: Dataset configuration file compatible with YOLOv8/v11 training.

How to Generate the Full Dataset

Clone the repository and run the generator script. The script uses multiprocessing to generate data quickly.

Recommended Command (High Quality)

# Generates 100k training images at 256x256 resolution
python -m scripts.data.synthetic_data_generator \
  --n-train 100000 \
  --n-val 10000 \
  --n-test 10000 \
  --imgsz 256 \
  --dataset-name lol_minimap_synthetic \
  --viewport-sim