Cornea

Cornea is a server that can take multiple video streams and perform facial recognition on them, returning the face match and the confidence value.

Currently Cornea is only capable of receiving data frame by frame by use of a REST API.

Example usage

In this example we send a base 64 encoded frame to the API using the Python Requests module and receive back a confidence rating between 0 and 1.

import requests

res = requests.post(
    'localhost:8000/detect_frame', json={
        "frame": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAIBAQEBAQIBAQECAgI..."
    }
)

print(res)

Which could return an output similar to:

>>> {
        'tag': 1,
        'confidence': 0.5711379345492313,
        'position': {
            'x': 555,
            'y': 179,
            'w': 177,
            'h': 177
        }
    }

Installation

To install and use you will need a working installation of PostgreSQL on your system.

Once setup run the following commands

CREATE ROLE cornea WITH LOGIN PASSWORD '<your_password>';
CREATE DATABASE cornea OWNER cornea;

Then, go ahead and install and run Cornea:

$ git clone https://github.com/euab/cornea.git

$ cd cornea

$ python3 -m cornea --run

On running the program for the first time, you will be prompted to configure the database connection according to how the database was created above. Edit the configuration stored in config.yml accordingly. An example is provided:

database:
    postgres:
        database: "cornea"
        user: "cornea"
        password: "youshallnotpass"
        host: "127.0.0.1"
        port: 5432

Ingesting training data

Currently, as we haven't implemented few-shot learning, a decent supply of training data is required to ensure accuracy. The implementation of few-shot or even one-shot learning methods is planned in the future. There is a schema for ingesting training data which is that a folder of training data will contain the training data for one person. You also need to supply their tag for the moment. So, a command to ingest training data for one person is:

# python3 -m cornea --ingest ./my_training_data/person1/ --tag 1

Training

To train a model, ensure that you have a populated database of valid training data and run:

$ python3 -m cornea --train

Models are outputted in YAML format.