myFace

This project finally shall include a trainable convolutional network to recognize familiar faces (distinguish them).

08.03.'18 to do: renew image database and implement yet planned tag_v2.0 get diff to tag_v1.4

Part 1 - Video

• capture - how to collect ?

  any video? write android app?

• name/tag

• crop

  A pretrained face-finder might become handy. Use 'Haar-Cascade-Classifier' or 'openface', or somekind of googles 'inception'

Openface would be really cool to use - mighty face detection. Eventually, it's useful to analyze each frame already here and just store tagged images of cropped detected faces. (opencv's annotate does that manually and writes face positions in images to a file.) Openface is probably able to cover this whole project by itself, but for now I failed to use it in anyway. So a Cascade Classifier might be sufficient (not for lateral faces without an extra classifier), what i'd like to cover in some way.

Part 2 - CNN

• load videofiles (at least two different classes to distinguish) and grab tagged frames

• train CNN

• Hyperparameters:

  • layers
  • nodes
  • filter/kernel size
  • stride
  • pooling

• save *.ckpt

Part 3 - Model Usage

• load model *.ckpt and feed in image to determine

show video or just return a string?