This repository provides details about image data:

• downloaded from Steam,
• filtered following parts of the procedure in download-steam-banners-data.

A dataset, called Steam-OneFace, is shared in this section.

Data downloaded from Steam consists of:

• PNG logos with transparency,
• JPG vertical banners.

Filtered image data is shared on Google Drive.

The up-to-date list of appIDs for which I have tried to download image data is available in:

data/download-queries/app_ids.txt. Most of these ~48k appIDs do not feature any image data.

Here is an example of a vertical banner:

Here is an example of a logo:

The filtering consists in the removal of:

• blank images, i.e. images for which grayscale image intensity extrema are equal,
• images of uncommon resolution:
  • anything but 640x360 for logos,
  • anything but 300x450 for vertical banners,
• images with uncommon bands:
  • anything but RGBA for logos,
  • anything but RGB for banners,
• images for which the bounding box of the non-zero regions covers:
  • 100% of the image space for transparent logos,
  • strictly less than 100% image space for banners, which happens with vignetting.

Suggestions of filtering include:

• removal of duplicate images with imagededup,
• filtering based on the number of detected faces, as in steam-face-detection.

The enforcement of such filtering is left to the reader. Otherwise, it would be difficult to keep filtered data up-to-date.

The notebook build_steam_oneface_dataset.ipynb shows an application of the filters suggested above.

There are three options for the face detector:

• face-alignment
• retinaface_pytorch
• dlib

This allows to build a dataset, called Steam-OneFace, of 1688 images which should all feature exactly one face.

This dataset is shared on Google Drive in:

• the original resolution (300x450): steam-oneface-hr.tar.gz (94 MB)
• a lower resolution (256x256): steam-oneface-lr.tar.gz (52 MB)

To use this dataset on Google Colab, run the following:

!gdown --id 1QptHrW9vloTtP--YJsxMY8PZWI2D8NJt
!tar xf steam-oneface-lr.tar.gz

import glob
from pathlib import Path

file_names = glob.glob('steam-oneface-lr/*.jpg')
app_ids = [int(Path(fname).stem) for fname in file_names]

The dataset consists of 2472 images, shared in:

• the original resolution (300x450): steam-oneface-hr_with_retinaface.tar.gz (133 MB)
• a lower resolution (256x256): steam-oneface-lr_with_retinaface.tar.gz (74 MB)

To use this dataset on Google Colab, run the following:

!gdown --id 1-0Nk7H6Cn3Nt60EdHG_NWSA8ohi2oBqr
!tar xf steam-oneface-lr_with_retinaface.tar.gz

The dataset consists of 305 images, shared in:

• the original resolution (300x450): steam-oneface-hr_with_dlib.tar.gz (16 MB)
• a lower resolution (256x256): steam-oneface-lr_with_dlib.tar.gz (9 MB)

To use this dataset on Google Colab, run the following:

!gdown --id 1-4RIn9G9Bee2JZ1bK1gkkgkLocHuWJJ4
!tar xf steam-oneface-lr_with_dlib.tar.gz

The notebook trim_steam_oneface_dataset.ipynb trims the dataset by intersecting the results of different detectors.

The trimmed datasets are:

• Steam-OneFace-small:
  • 993 images,
  • obtained with modules face_alignment and retinaface,

To use this dataset on Google Colab, run the following:

!gdown --id 1-1V5fDhPo75iDtAbrD18rppV-lf51bPW
!tar xf steam-oneface-small-lr.tar.gz

• Steam-OneFace-tiny:
  • 168 images,
  • obtained with modules dlib, face_alignment and retinaface.

To use this dataset on Google Colab, run the following:

!gdown --id 1-2sCVgBUmu6LFug1pzBfmL8zNFFBq27F
!tar xf steam-oneface-tiny-lr.tar.gz

• 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
• To detect faces: detect_faces_on_steam_banners.ipynb in woctezuma/steam-face-detection