ReviewFaceRecognition

Review Face Recognition packages such as ClarifAI, face_recognition (dlib) and Microsoft Azure

Motivation: For my research, I was tasked to evaluate exisiting face recognition packages and compare their performances. Although, the goal was to use Faster RCNN for both face identification and recognition, Faster RCNN model we used was trained only to identify. As a result, we needed to find package that would classify already identified faces.

Goal Find a package that would recognize faces already identified (and cropped) by Faster RCNN as 'face'.

Step 1 Faster RCNN

• My tutorial on how to install Faster RCNN can be found here
• Extract faces from pictures and crop them using Faster RCNN and OpenCV.

NOTE All the images passed through Testing APIs are classified as 'face' by Faster RCNN.

Step 2 Installation of Testing APIs

• Microsoft Azure

  • Hardest and least intuitive API (among these APIs) to install. Perhaps it was my inexperience, but I found online modules and guides provided by Microsoft to be unclear, and I couldn't find much information on th internet on how to install/deploy the package.
  • Note: when using Free Tier you are limited to 10,000 requests per month AND 20 calls per minute. In other words, you will be able to evaluate only 20 images per minute.
  • Installation guides can be found here
    • Note: Suppose you want to classify faces, group images or find similarities between them. First, you will have to detect faces in the images and then you will have to separately run another code that will call . All in all, you will have to run two separate calls. For example, in this guide, I only run face detection, which means that Azure only identified whether faces existed in the following images or not. If I wanted to group images together, I would have had to run another code (or modify my existing one) to make grouping happen. In sum, it would have been two different operations. More information can be found here
  • You can use my code named azure_demo.py
    • Values to change: uri_base - you need to provide your region when you created account. Also, you will need to add subscription_key
    • Input: inputDir - is a location of the folder that needs to be evaluated
    • Output: identified.txt - is txt file will print out image name and unique ID associate with image. I provided my file for illustrative purposes. Whenever you see image name and ID next to it, it means that Azure was able to identify the image as a face. It doesn't mean that Azure was able to recognize the image.
      • Note 2: Whenever Azure produces an output for each evaluated image, it produces uniquie ID ca810040-6727-4500-895d-f2258bbebf83 which I found to be somewhat inconvenient. So, in order to see your result, you must match it with the image you evaluated. I stored result in the following (key-value) format: imageName: ca810040-6727-4500-895d-f2258bbebf83 , where imageName was treated as key.

• Clarifai

  • Perhaps the easiest package to install and employ from the following APIs. First, create account and get you API Key. Second, open terminal and type the following command: pip install clarifai
  • Clarifai provides many different models such as apparel, food, demographics, and face detection. However, for the purpose of this post, I used face detection
  • Installation guide can be found here
  • You can use my code (written in python) located in clarifai folder that are titled:
    • For demo.py if you want to test a single image.
      • Change API key : app
      • Input: image path (string) to variable fName
      • Output: prints out string in JSON format
    • For celebrity_test.py if you want to test folder.
      • Change API key : change variable app
      • Input: directory inputDir
      • Output: 2 txt files named vectors.txt and identified.txt
        • identified.txt is a text file that has the following format: imageName : identified face or not
        • vector.txt is a text file where you get imageLocationPath : either 1024 vector representation or [] if no image

• face_recognition

  • My tutorial on how to install face_recognition package can be found here
    • Note: installing package requires only 1 line of code. However, you must get DLIB to work properly, which is usually the hardest part.
  • Peculiar thing about this package is that it can take only 1 image as a training sample (i.e. you can't provide more than 1 training sample for each face you want to identify/recognize).
  • Based on OpenFace and Dlib, face_recognition package is easy to implement and has 99% accuracy rate on LFW benchmark. Written by Adam Geitgey who also has blog posts with tutorials

Note Azure allows only 10,000 calls/month, and Clarifai 5,000 calls/month, whereas face_recognition is open source

Step 3 Testing

• All the images given to Testing APIs were already cropped images from Faster RCNN. The images (faces) given were from TV show 'Friends' Episode 2 Season 6
• Although all the images passed where identified as 'face' by Faster RCNN, Testing packages had different results

Face Recognition API	Identified	Total	% (Identfied)
Microsoft Azure	6038	10,725	53%
Clarifai	2765	4558	60%
face_recognition	1959	4558	43%

• Using the following table to compare 3 APIs is not quiet accurate. The reason is that the total amount of images evaluated wasn't equal. As a result, in order to compare three APIs, you have to evaluate same number of images in all three.
• I picked 4558 as max images evaluated because I ran out of free API calls on Clarifai

Face Recognition API	Identified	Total	%
Azure	2066	4558	45%
Clarifai	2765	4558	60%
face_recognition	1959	4558	43%

Verdict Considering that face_recognition is free, it performs quiet well compared to Azure. In fact, if you only need to recognize and identify faces, it is not worth using Microsoft Azure. However, note that Azure can also give you age, gender, hair color, and items in image (like glasses or hat). Clarifai performs quiet well compared two above mentioned packages. Nevertheless, Clarifai was able to identify only 60% of all images Faster RCNN identified as face. If you consider that Faster RCNN identified roughly 10% as false positives, there is still 30% that Clarifai was unable to identify.