Awesome

Steam Face Detection

<img alt="Input" src="https://raw.githubusercontent.com/wiki/woctezuma/steam-face-detection/img/cover-input.jpg" width="375"> -> <img alt="Output" src="https://raw.githubusercontent.com/wiki/woctezuma/steam-face-detection/img/cover-output.jpg" width="375">

This repository contains Python code to detect faces on Steam store/library images.

Usage

• Acquire the data, e.g. as a snapshot called 256x256.zip in one of my data repositories (here or there),
• Run detect_faces_on_steam_banners.ipynb to detect (and count) faces.

NB: In order to try and adjust parameters, I suggest that you experiment with benchmark_face_detection.ipynb.

Data

The dataset consists of 14k vertical images, resized from 300x450 to 256x256 resolution, used by the Steam library.

Images were downloaded with download_steam_banners.ipynb.

Images were then filtered (duplicates, outliers, etc.) with remove_duplicates.ipynb.

Recommendations

I have used 3 tools for face detection:

• dlib, which I used in another one of my projects,
• face-alignment,
• retinaface, available on PyPI.

Overall, I would recommend to use face-alignment over dlib as:

• it detects more faces on Steam images, which feature difficult faces (hand-drawings, anime, 3D models, etc.),
• it is about 3.5 times faster.

RetinaFace can detect even more faces, but it is slower than face-alignment.

Results

Overall, most images do not feature any detected face. Among images with faces, most images feature a single face. The more detected faces, the fewer images.

Here are a few face detection results:

Keep in mind that the algorithm is not foolproof!

References

• To download images: download_steam_banners.ipynb in woctezuma/google-colab
• To filter out duplicates, etc.:
  • for PNG logos: remove_duplicates.ipynb in woctezuma/google-colab
  • for JPG banners: remove_duplicates.ipynb in woctezuma/steam-stylegan2-ada
• Python packages for face detection:
  • face-alignment
  • retinaface
• Data stored on Google Drive:
  • filtered images: woctezuma/download-steam-banners-data
  • filtered images, with less aggressive filters: woctezuma/steam-filtered-image-data
  • the Steam-OneFace dataset: Steam-OneFace