Scribd
Upload
130 views12 pages

Face Detection Poc: Using Opencv and Other Freely Available Libraries

This document evaluates different libraries and algorithms for face detection using OpenCV, Dlib, and OpenFace. It tests Haar cascade, DNN, HoG, and MMoD algorithms on CCTV footage. DNN achieved the best results but only 6 fps. Increased image size improved detection but reduced speed further. While Haar cascade had many false positives, DNN handled variations in orientation, occlusion, and scale best. The document recommends DNN and notes other useful OpenCV functions.

Uploaded by

Gareth Alldread
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
130 views12 pages

Face Detection Poc: Using Opencv and Other Freely Available Libraries

This document evaluates different libraries and algorithms for face detection using OpenCV, Dlib, and OpenFace. It tests Haar cascade, DNN, HoG, and MMoD algorithms on CCTV footage. DNN achieved the best results but only 6 fps. Increased image size improved detection but reduced speed further. While Haar cascade had many false positives, DNN handled variations in orientation, occlusion, and scale best. The document recommends DNN and notes other useful OpenCV functions.

Uploaded by

Gareth Alldread
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 12

Face Detection POC

Using OpenCV and Other Freely Available Libraries


Libraries

 OpenCV
 Open source computer vision and machine learning software library.
 Dlib
 Open source C++ toolkit containing machine learning algorithms.
 OpenFace
 Implementation of face recognition with deep neural networks.
 Discounted as Linux focused build scripts and beyond scope of POC.
 May be worth revisiting in the future – quick overview on last slide.
 All Support GPU Processing using OpenCL.
Test Environment

 Test app written in C++ (Also both libraries have excellent Python support)
 Initially Used PBE as Frame Server.
 No Success – Probably Due to Colour Space or Frame Format Incompatibility.
 Created MP4 files from SynAV2 footage for processing.
Algorithms

 OpenCV Haar Cascade


 State-of-the-art in Face Detection in 2001.
 Uses a cascade of elimination based
on simple detection of thousands of
feature types.

 OpenCV DNN (Deep Neural Network)


 Uses a pre-trained model included as part of OpenCV.
 The model was trained using images available from the web, but the source is
not disclosed…
Algorithms Continued
 Dlib HoG (Histogram of Oriented Gradients)
 Extracts edge gradients from each square on a grid.
 Matches a histogram to a pre-trained pattern (provided as part of Dlib).

 Dlib MMoD (Maximum-Margin Object Detector)


 Neural net trained on images from various datasets (ImageNet, PASCAL VOC, VGG,
WIDER, Face Scrub) containing 7220 images – freely available.
Results

 Original sample code expected 300p image


 Fast (25fps) but VERY poor results from typical CCTV FOV.
 Worked well with webcam footage.
 After Further Research – Recommendation was to increased to 900p
images.
 Much better results - but much slower - around 5 fps.
 GPU reduced CPU load dramatically but did not improve the frame rate.
Results: OpenCV Haar Cascade
 Lots of false positives, doesn’t work on non-frontal images, doesn’t work with occlusion.
Results: Dlib

 HoG (Histogram of Oriented Gradients)


 Low incidence of false positives.
 However - LOTS of missed faces.
 Very sensitive to scale.

 MMoD (Maximum-Margin Object Detector)


 No success finding faces at all in my testing.
 VERY low framerate (< 1fps)
Results: Open CV DNN (Deep Learning)
 Best results by far. Handles different orientations, occlusion and scale variations.
 However only capable of 6 fps.
Results: Open CV DNN (Deep Learning)
 Second example:
Other OpenCV Functionality of Value

 Motion Detection / Background Removal.

 Face Recognition using a Deep Learning Model.

 Excellent Video File Reading/Writing Capabilities.

 General Video/Image Manipulations inc. Overlays.


Open Face
 Detects faces using OpenCV.
 Transform the face for the neural network.
 Use a deep neural network to represent the face on a 128-dimensional unit
hypersphere.

 Calculate 128-dimension distance from pre-trained faces to classify.


 Neural net has been trained to minimize distance for same person, maximize
distance for different people. (This is the technique that Apple is believed to
use)

You might also like