AAM/POSIT face detector by Damian Leśniak, 6th Jun 2015

GENERAL INFORMATION:

This software was strongly inspired by: "Mastering OpenCV with Practical Computer Vision Projects", Chapter 7 https://github.com/MasteringOpenCV/code/tree/master/Chapter7_HeadPoseEstimation by Daniel Lélis Baggio

Author of the original AAM method: Tim Cootes (Professor of Computer Vision)

Creator ot the 3D face model used in this software: Pedro A. D. Martins (PhD) http://aifi.isr.uc.pt/Downloads/OpenGL/glAnthropometric3DModel.cpp

Comprehensive description of the AAM algorithm: P. Martins, "Active Appearance Models for Facial Expression Recognition and Monocular Head Pose Estimation"

HOW TO INSTALL:

cd path_to_project_directory cmake . make

BINARIES:

get_data: use the mouse to drag points across the face snapshot 'c' - start/pause webcam 's' - save changes (warning! all unsaved changes are lost if you unpause the webcam or use keys '>', '<') 'i' / 'u' - rotate shape (warning! this action, if performed repeatedly, results in distiortions of points' coordinates) 'j' / 'k' / 'l' / 'h' - move shape '>' / '<' - go to next/previous shapshot '+' - make a copy of current snapshot and append it to the list of images 'e' - equalize the snapshot (not recommended) 'z' - show/hide edges 'x' - show/hide vertices

train: creates AAM and stores it in the data directory

test_model: you may check out the parameters learned by AAM and see if POSIT is working correctly

final: sit back and enjoy AAM/POSIT struggling to recognize you

HOW TO CREATE A WORKING AAM/POSIT:

Option 1:

• grow a beard
• run the 'train' executable

Option 2:

• go to the 'images' directory
• remove all but one images (leave '1.jpg')
• go to the 'data' directory
• edit the 'shapes.txt' file
  • change the first line to '1 116'
  • don't modify the second line
  • delete all the following lines
• run the 'get_data' executable to modify the first image and create at least 20 new images
• run the 'train' executable
• run the 'test_model' executable to check if everything went according to the plan

Option 3 (advanced, not tested, might require additional changes in the source code):

• create your own 3D shape, sort the vertices
• go to the 'data' directory and modify the following files (the first line should always be: 'number_of_rows number_of_cols'
  • model3D.txt - store point coordinates in each row
  • reference_shape.txt - project the 3D model into 2D space, store the coordinates in one row: 'x1 y1 x2 y2 ... xn yn'
  • shapes.txt - replace with a copy of 'reference_shape.txt'
  • triangles.txt - triangulate your 3D model, store each triangle (described by its vertices' numbers) in one row
• edit the 'main.cpp' source file, delete (or replace) the code using 'face_cascade' to estimate position of the face
• go to the 'images' directory
• remove all but one images (leave '1.jpg')
• run the 'get_data' executable to modify the first image and create new images
• run the 'train' executable
• run the 'test_model' executable to check if everything went according to the plan

TUNABLE PARAMETERS:

• in 'main.cpp': MAX_ALLOWED_ENERGY - used to check if AAM approximation is good enough