Greety recognizes persons appearing in front of the camera and greets them with a joyful message. Just train greety on faces you want it to recognize, attach a camera and let the magic begin.

Start by cloning the project with the openface subproject git clone --recursive https://github.com/presight/greety

Setup openface as described at http://cmusatyalab.github.io/openface/setup/

Get dlib face landmarks, which are used for detecting faces ./openface/models/dlib/get-models.sh

The first step is to collect images of faces you want greety to recognize in a folder, let's call it {dir}.

If you have a set of uncategorized images and want to extract as many faces as possible, this command will extract and align all face images found in the images and put them in generated/aligned/

python ./align_dlib_multiple.py {dir} align outerEyesAndNose generated/aligned --size 96

After the aligned images have been generated you have to move the images to sub folders named after the desired labels, for example {dir}/person1.

Put all images in sub directories for each person, for example {dir}/person1 etc, then run the following command:

python ./openface/util/align_dlib.py {dir} align outerEyesAndNose generated/aligned --size 96

Run feature_saver.py and all found faces will be saved in generated/unknown. Remove bad images and move the remaining images to sub directories in {dir} corresponding to their labels.

First remove the cache from eventual previous runs rm generated/aligned/cache.t7

Generate face representations ./openface/batch-represent/main.lua -outDir ./generated -data ./generated/aligned/

Train greety classificator, optionaly with a specified conf file python train.py or python train.py default.conf

Run greety, optionaly with a specified conf file python greety.py or python greety.py default.conf

Download a set of unknown images not included in the dataset to learn, for example a subset of http://vis-www.cs.umass.edu/lfw/. Generate ./generated/unknown.npy from the images in {lfw_directory}. The resulting file will contain face embeddings that will represent the unknown faces. ./openface/demos/web/create-unknown-vectors.py --outputFile ./generated/unknown.npy --dlibFacePredictor ./openface/models/dlib/shape_predictor_68_face_landmarks.dat --model ./openface/models/openface/nn4.small2.v1.t7 {lfw_directory}`

Also make sure unknown_reps in the config point to the generated unknown.npy file

To get text to voice synthesizing working you can install espeak or marytts, or implement support for a lib of your choosing.

So far openface hasn't reached a good decision on how to classify unknown faces, and thus the current system with DBN generates some false positives.