A hurricane side project that spectates League for you

What it does DeepLeague ingests 1080p League of Legends video—from a YouTube URL, a local MP4, or a folder of images—and isolates the minimap so a neural network can draw bounding boxes around champions. It wraps this in a small CLI tool, test_deep_league.py, and relies on a custom fork of YAD2K with pre-trained weights to do the heavy lifting. The end result is a directory of annotated frames showing exactly where each player is standing on the map.

The interesting bit The bulk of the code was written in five days during a hurricane, and the author freely admits the project is more of a spirited prototype than a polished library. The hidden gem is the 100,000-image dataset—already carved into training, testing, and validation splits and served as a 30GB archive—because labeled esports data is usually harder to find than the model code itself.

Key highlights

• Accepts three input sources: YouTube links, local MP4 files, or directories of images.
• Built on a custom fork of YAD2K with downloadable pre-trained weights, so you can run inference without training from scratch.
• Bundles a 100,000-frame labeled dataset of minimap crops and bounding boxes, split 80/17.5/2.5 for train/test/val.
• Hard-locked to 1080p input; the crop logic and model are untested on any other resolution.
• Outputs annotated frames to an output folder for further analysis.

Caveats

• The author warns of bugs and inefficiencies, and says anyone wanting to modify core code is “mostly on your own.”
• Model setup is system-dependent; the README notes that weight conversion often throws hardware-specific errors you are expected to Google.
• Only 1080p VODs are supported, so off-resolution footage will break the minimap crop.

Verdict Worth a look if you are doing esports computer-vision research and need a minimap-specific dataset or a hackable YOLO pipeline to fork. Skip it if you want a refined, resolution-agnostic tool with stable APIs and hand-holding documentation.