Skip to content

Latest commit

 

History

History
83 lines (61 loc) · 3.84 KB

README.md

File metadata and controls

83 lines (61 loc) · 3.84 KB
Raw

My Final Year Project - A Comparison of Facial Recognition Implementations

This is a refactor of my existing work to encapsulate everything into one modular project. The goal of this project is compare deep learning techniques against more traditional approaches in the task of facial recognition. I will be comparing performance and efficiency in two key areas:

  1. Quality of Facial Embeddings
  2. Quality of Classifications

For 1), I will be comparing FaceNet embeddings against HOG feature vectors from OpenCV. For 2), I will be comparing a Neural classifier against an SVM

These files will allow you to choose a form of getting embeddings, choose a classifier and it will then run and measure the performance.

This is still a work in progress.

Implemented Features

  1. Dataset Parsing
  2. Obtaining HOG Feature Vectors
  3. SVM Classifier
  4. Obtaining FaceNet Embeddings
  5. K Nearest Neighbours Classifier
  6. Save trained SVM
  7. Adam Geitgey's Face Recognition Encoding

Features still to be Implemented

  1. Neural Classifier
  2. Improved Performance Measuring

Prerequisites

This requires Python 3.X

A working Tensorflow environment is required. Install instructions can be found here.

Tested using the LFW Dataset. You can download this here.

You will require the FaceNet code for the FaceNet embedding part of this project. You can download this by executing: git clone https://github.com/davidsandberg/facenet Following that you will need to add it to your python path. export PYTHONPATH=[...]/facenet/src (Ensure to adjust to the path where you have downloaded FaceNet to.) You can add this line to the bottom of your ~/.bashrc file to ensure that it is run every time you open a terminal.

Download a vesion of a pre trained FaceNet model. The one I used for my testing can be found here

Some packages need to be installed from the Ubuntu repos. I have made a script to cover this; just run the following commands: chmod +x requirements.sh sudo ./requirements.sh

Python requirements can be found and installed from requirements.txt pip3 install -r requirements.txt

Parameters

Run the project by executing python3 project.py with the following arguments:

  • --embedding Select a valid embedding implementation. (Valid options are "facenet", "hog_opencv", "hog_scikit" and "dlib"[for adam geitgeys encodings])
  • --classifier Select a valid classifier implementation. (Valid options are "svm" or "knn")
  • --dataset Provide full path to the dataset. (Only tested with lfw)
  • --min_nrof_images_per_class Provide minimum number of images required for a class to be included
  • --num_test_images_per_class Provide the number of test images required per class
  • --mdlpath Provide full path to the tensorflow facenet model (.pb file)
  • --gpu_memory_fraction Upper bound on the amount of GPU memory that will be used by the process
  • --use_trained_svm Path to a pre-trained, saved svm (.pkl file)

Note: If the "--use_trained_svm" argument is not used, the program will automatically save the new svm to the current directory as a .pkl file. This .pkl file is what can be used again with the "--use_trained_svm" tag

Example

python project.py --embedding facenet --classifier svm --dataset /<dir_to_dataset>/datasets/lfw/ --min_nrof_images_per_class 10 --num_test_images_per_class 3 --mdlpath /<dir_to_model>/models/20170514-110547/20170512-110547.pb

Results To Date

4 Jan 2018

SVM Neural Net KNN
Facenet 68.8%-22secs - 99.9%-.025mins
OpenCV HOG 66.1%-25hours - 22.2%-20mins
DLib 100%-15mins - -