Module 3 Assignment

Module 3

Assignment

Table of Contents

Train custom Facial Landmark Detector model                               2

Improve Facial Landmarks Stabilization                                          3

Use OpenCV’s face detector                                                             3

Assignment 3-1

1. Train custom Facial Landmark Detector model

Train a facial landmark detector model with 21 points or with any number of points you desire.

Example output for 21-points facial landmark detector model.

I don’t want to be original, so I will do it exactly as they asked, and in the same position.

Procedure:

I will read _bv70.txt files as input, pick then 21 landmarks and save them as _bv21.txt, one file for each image.

I modified the code generate33Points.py and save it as generate21Points.py. This program outputs two files sets in the directory  facial_landmark_data>datasets>afw

Aaaa_bv21.txt

……………………..

Yyyy_bv21.txt

Each one has 42 numbers for 21 coordinates of 21 landmarks.

I run      python  drawRectLandmarks.py   d:\cv4faces\facial_landmark_data\   21

With that I revised image  output files “original”, “mirror”

I run     python  createTrainTextXml.py   d:\cv4faces\facial_landmark_data\   21

I run     python   trainFLD.py    d:\cv4faces\facial_landmark_data\   21

It run for 50 minutes and output on the consola

Training accuracy:  8.2868806

Testing accuracy: 11.15939011140127

And the file shape_predictor_21_face_landmarks.dat

I got these images with the program facialLandmarkDetector.cpp modified.

21 landmarks on Hillary-Clinton and 21 landmarks on Clinton-Hillary.

21 landmarks on Donald-Trump and 21 landmarks on Trump-Donald.

2. Improve Facial Landmarks Stabilization

Try to improve the code on facial landmark stabilization. Can you make the landmarks more stable?

Hint : Some landmarks are more unstable than others. Can we treat them differently?

3. Use OpenCV’s face detector

Use OpenCV instead of Dlib for face detection and pass this as an input to facial landmark detection. Do you see a difference in accuracy and performance of face detection + landmark detection code?

The rectangles of the faces detected are different in size and in the position output from  the face detector of Dlib, and the OpenCV face detectors Haar and Lbp.

Dlib seems to be more precise, in the family photo I got “false positives” for Haar and Lbp. In the boy photo I got “false positives” with Lbp.

I detected small differences in position of landmarks in some images, as they depend on the rectangle passed to Dlib. (for example in Hillary lips).

Next week more on this course.