These tools are useful for working with face detection using OpenCV. They are writtin in Python 3 but should work with Python 2 relatively easily. I plan to continue to work and improve this collection, but would appreciate any feedback.

Assumes you have the following installed:

• Python 3
• OpenCV
• NumPy

These scripts provide different functionality or use different algorithms:

• eval_face_training.py - Uses Deep Neural Networks to evaluate your face training dataset. It counts a good face image an image with a single face with a confidence factor that you set (defaults to 50%). The an example and parameters are as follows: python eval_face_training.py -d training_images -p deploy.prototxt.txt -m facedetection.caffemodel
  • --dir (-d) [path] : The path to the training data to look at
  • --prototxt (-p) [file]: Path to the Caffe 'deploy' prototxt file
  • --model (-m) [file] : Path to the Caffe pre-trained model (ending .caffemodel)
  • --confidence (-c) int : Minimum confidence percent to filter out weak face detections, 1.0 max
  • --multifaces (-f) : If set and there is multiple faces, will treat highest one as the good one
  • --display_problems : Will show a window of each failed image that has 0 or >1 faces in it (faces in a red rectangle)
  • --verbose (-f) : Turn on verbose output (each file parsed, etc)
• eval_face_training_cascade.py - Uses Haar Cascades to evaluate face training data (like eval_face_training.py). The example and parameters are as follows: python eval_face_training_cascade.py -d training_images -p haarcascade_frontalface_default.xml
  • --dir (-d) [path] : The path to the training data to look at
  • --cascade (-c) [file] : Path to the face Haar Cascade file to use
  • --display_problems : Will show a window of each failed image that has 0 or >1 faces in it (faces in a red rectangle)
  • --verbose (-f) : Turn on verbose output (each file parsed, etc)

I like to use the DNN based face data evaluators for finding good face training data of individuals. For me to set that up:

1. Download images that I want to train with into a local directory
2. Run the python eval_face_trainin.py file against that directory and it will tell me overall how many images it thinks are "good" (one face meeting the confidence factor): python eval_face_training.py -d training_images -p deploy.prototxt.txt -m res10_300x300_ssd_iter_140000.caffemodel
3. The output will be something like this:

Evaluating directory: training_images
   Total dirs: 2
   Total files: 24
   Total faces: 23
   Total images: 24
   Total good images: 21
   Overall score: 87.5%
   Execution time: 1.54642605782 seconds (0.0644344190756 sec avg per file)

4. I can see overall most of the images were good (yeah!) but I may want to throw some out. If I specify verbose (-v) then it will tell me specifically for each file how it did:

   ...
loading file: training_images/s2/9.jpg
   1 faces
loading file: training_images/s2/12.jpg
   0 faces
loading file: training_images/s2/11.jpg
   1 faces
loading file: training_images/s2/10.jpg
   ...

I can now remove/replace those files if I wish, or I can turn on '--display_problems' flag and run again and it will show me the images. This is also useful if I get an image that has more then 1 face in it.

• Chris Whitten - Initial work - Github

This project is licensed under the MIT License - see the LICENSE.md file for details

Inspiration:

• Siraj Raval - For dispensing incredible knowledge on computer vision and machine learning
• Adrian at PyImageSearch - For great learning material on how to use OpenCV and use DNNs