wikifaces

A downloader for named images containing faces from Wiki servers.


License
Other
Install
pip install wikifaces==1.0.8

Documentation

Wiki Faces:

License: MIT License

Figure 1: Joko Widodo's Wikipedia page, which includes am image of his face. The cropped image on the right is download into a directory named "Joko_Widodo."

TLDR

This project downloads images from a Wiki that include human faces. Specifically, images that are associated with certain wikipedia categories.

Installation

Pip Installation Procedure:

From PIP:

pip install wikifaces 

From Repo:

git clone git@github.com:tford9/Wiki-Faces-Downloader.git
cd Wiki-Faces-Downloader
python setup.py
pip install wikifaces

Usage

Command-Line Example

python downloader -i "indonesian engineers" -o ../data/ -d

Package Example

from wikifaces.downloader import WikiFace

wikiface_obj = WikiFace()
wikiface_obj.download(categories=['facebook'], depth=2, output_location='../data/')

The following structure is output:

  • facebook
    • cached_1_people_pages_d2.pkl
    • cached_pages_d2.pkl
    • alan_rushbridger
      • Alan_Rusbridger_01.jpg-p0.jpg
    • ...
    • mark_zuckerberg
      • MarkZuckerbergcrop.jpg-p1.jpg
    • ...

The process is carried out as follows:

  1. Given a category from a Wiki, collect n pages that contain the same category as well as at least one category containing "people" in the title.
  2. With those pages, crawl across their included categories and collect y pages that contain those categories as well as at least one "people" category.
  3. Given the collected Wiki pages, download the primary image from the page and determine if it is a human face using light facial detection.
  4. We capture all images from the wiki that contain the name of the page (if it's a person then the filename contains their name),
  5. Using the captured name and images, we create a dataset for that face.

TODOs:

  1. Currently, a part of this process uses a recursive call structure to get all related pages; there may be a way to linearize, or parallelize this.
  2. Currently, we are only pulling images contain the person's name in the title and only have one visible face in the image. All other images are not considered. A voting system should be added to get the most represented faces across multiple images.