Chiebot-Mirror
/
opencv_contrib
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
Repository for OpenCV's extra modules
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
2962 Commits
5 Branches
55 Tags
401 MiB
 
 
 
 
 
 
Tag: Branch: Tree: 4df17e67bd
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '4df17e67bd'
${ noResults }
opencv_contrib/modules/face/src/facemarkAAM.cpp

1220 lines
41 KiB
Raw Blame History

/*
By downloading, copying, installing or using the software you agree to this
license. If you do not agree to this license, do not download, install,
copy or use the software.
License Agreement
For Open Source Computer Vision Library
(3-clause BSD License)
Copyright (C) 2013, OpenCV Foundation, all rights reserved.
Third party copyrights are property of their respective owners.
Redistribution and use in source and binary forms, with or without modification,
are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the names of the copyright holders nor the names of the contributors
may be used to endorse or promote products derived from this software
without specific prior written permission.
This software is provided by the copyright holders and contributors "as is" and
any express or implied warranties, including, but not limited to, the implied
warranties of merchantability and fitness for a particular purpose are
disclaimed. In no event shall copyright holders or contributors be liable for
any direct, indirect, incidental, special, exemplary, or consequential damages
(including, but not limited to, procurement of substitute goods or services;
loss of use, data, or profits; or business interruption) however caused
and on any theory of liability, whether in contract, strict liability,
or tort (including negligence or otherwise) arising in any way out of
the use of this software, even if advised of the possibility of such damage.
This file was part of GSoC Project: Facemark API for OpenCV
Final report: https://gist.github.com/kurnianggoro/74de9121e122ad0bd825176751d47ecc
Student: Laksono Kurnianggoro
Mentor: Delia Passalacqua
*/
#include "precomp.hpp"
#include "opencv2/face.hpp"
namespace cv {
namespace face {
/*
* Parameters
*/
FacemarkAAM::Params::Params(){
model_filename = "";
m = 200;
n = 10;
n_iter = 50;
verbose = true;
save_model = true;
scales.push_back(1.0);
max_m = 550;
max_n = 136;
texture_max_m = 145;
}
FacemarkAAM::Config::Config(Mat rot, Point2f trans, float scaling,int scale_id){
R = rot.clone();
t = trans;
scale = scaling;
model_scale_idx = scale_id;
}
void FacemarkAAM::Params::read( const cv::FileNode& fn ){
*this = FacemarkAAM::Params();
if (!fn["model_filename"].empty()) fn["model_filename"] >> model_filename;
if (!fn["m"].empty()) fn["m"] >> m;
if (!fn["n"].empty()) fn["n"] >> m;
if (!fn["n_iter"].empty()) fn["n_iter"] >> m;
if (!fn["verbose"].empty()) fn["verbose"] >> m;
if (!fn["max_m"].empty()) fn["max_m"] >> m;
if (!fn["max_n"].empty()) fn["max_n"] >> m;
if (!fn["texture_max_m"].empty()) fn["texture_max_m"] >> m;
if (!fn["scales"].empty()) fn["scales"] >> m;
}
void FacemarkAAM::Params::write( cv::FileStorage& fs ) const{
fs << "model_filename" << model_filename;
fs << "m" << m;
fs << "n" << n;
fs << "n_iter" << n_iter;
fs << "verbose" << verbose;
fs << "max_m" << verbose;
fs << "max_n" << verbose;
fs << "texture_max_m" << verbose;
fs << "scales" << verbose;
}
class FacemarkAAMImpl : public FacemarkAAM {
public:
FacemarkAAMImpl( const FacemarkAAM::Params &parameters = FacemarkAAM::Params() );
void read( const FileNode& /*fn*/ ) CV_OVERRIDE;
void write( FileStorage& /*fs*/ ) const CV_OVERRIDE;
void saveModel(String fs);
void loadModel(String fs) CV_OVERRIDE;
bool setFaceDetector(bool(*f)(InputArray , OutputArray, void * ), void* userData) CV_OVERRIDE;
bool getFaces(InputArray image, OutputArray faces) CV_OVERRIDE;
bool getData(void * items) CV_OVERRIDE;
bool fitConfig( InputArray image, const std::vector<Rect>& roi, std::vector<std::vector<Point2f> >& _landmarks, const std::vector<Config> &runtime_params ) CV_OVERRIDE;
protected:
bool fit( InputArray image, const std::vector<Rect>& faces, CV_OUT std::vector<std::vector<Point2f> >& landmarks ) CV_OVERRIDE;
bool fitImpl( const Mat image, std::vector<Point2f>& landmarks,const Mat R,const Point2f T,const float scale, const int sclIdx=0 );
bool addTrainingSample(InputArray image, InputArray landmarks) CV_OVERRIDE;
void training(void* parameters) CV_OVERRIDE;
Mat procrustes(std::vector<Point2f> , std::vector<Point2f> , Mat & , Scalar & , float & );
void calcMeanShape(std::vector<std::vector<Point2f> > ,std::vector<Point2f> & );
void procrustesAnalysis(std::vector<std::vector<Point2f> > , std::vector<std::vector<Point2f> > & , std::vector<Point2f> & );
inline Mat linearize(Mat );
inline Mat linearize(std::vector<Point2f> );
void getProjection(const Mat , Mat &, int );
void calcSimilarityEig(std::vector<Point2f> ,Mat , Mat & , Mat & );
Mat orthonormal(Mat );
void delaunay(std::vector<Point2f> , std::vector<Vec3i> & );
Mat createMask(std::vector<Point2f> , Rect );
Mat createTextureBase(std::vector<Point2f> , std::vector<Vec3i> , Rect , std::vector<std::vector<Point> > & );
Mat warpImage(const Mat ,const std::vector<Point2f> ,const std::vector<Point2f> ,
const std::vector<Vec3i> , const Rect , const std::vector<std::vector<Point> > );
template <class T>
Mat getFeature(const Mat , std::vector<int> map);
void createMaskMapping(const Mat mask, const Mat mask2, std::vector<int> & , std::vector<int> &, std::vector<int> &);
void warpUpdate(std::vector<Point2f> & shape, Mat delta, std::vector<Point2f> s0, Mat S, Mat Q, std::vector<Vec3i> triangles,std::vector<std::vector<int> > Tp);
Mat computeWarpParts(std::vector<Point2f> curr_shape,std::vector<Point2f> s0, Mat ds0, std::vector<Vec3i> triangles,std::vector<std::vector<int> > Tp);
void image_jacobian(const Mat gx, const Mat gy, const Mat Jx, const Mat Jy, Mat & G);
void gradient(const Mat M, Mat & gx, Mat & gy);
void createWarpJacobian(Mat S, Mat Q, std::vector<Vec3i> , Model::Texture & T, Mat & Wx_dp, Mat & Wy_dp, std::vector<std::vector<int> > & Tp);
std::vector<Mat> images;
std::vector<std::vector<Point2f> > facePoints;
FacemarkAAM::Params params;
FacemarkAAM::Model AAM;
FN_FaceDetector faceDetector;
void* faceDetectorData;
private:
bool isModelTrained;
};
/*
* Constructor
*/
Ptr<FacemarkAAM> FacemarkAAM::create(const FacemarkAAM::Params &parameters){
return Ptr<FacemarkAAMImpl>(new FacemarkAAMImpl(parameters));
}
/*
* Constructor
*/
Ptr<Facemark> createFacemarkAAM(){
FacemarkAAM::Params parameters;
return Ptr<FacemarkAAMImpl>(new FacemarkAAMImpl(parameters));
}
FacemarkAAMImpl::FacemarkAAMImpl( const FacemarkAAM::Params &parameters ) :
params( parameters ),
faceDetector(NULL), faceDetectorData(NULL)
{
isModelTrained = false;
}
void FacemarkAAMImpl::read( const cv::FileNode& fn ){
params.read( fn );
}
void FacemarkAAMImpl::write( cv::FileStorage& fs ) const {
params.write( fs );
}
bool FacemarkAAMImpl::setFaceDetector(bool(*f)(InputArray , OutputArray, void *), void* userData){
faceDetector = f;
faceDetectorData = userData;
return true;
}
bool FacemarkAAMImpl::getFaces(InputArray image, OutputArray faces)
{
if (!faceDetector)
return false;
return faceDetector(image, faces, faceDetectorData);
}
bool FacemarkAAMImpl::getData(void * items){
CV_Assert(items);
Data* data = (Data*)items;
data->s0 = AAM.s0;
return true;
}
bool FacemarkAAMImpl::addTrainingSample(InputArray image, InputArray landmarks){
// FIXIT
std::vector<Point2f> & _landmarks = *(std::vector<Point2f>*)landmarks.getObj();
images.push_back(image.getMat());
facePoints.push_back(_landmarks);
return true;
}
void FacemarkAAMImpl::training(void* parameters){
if(parameters!=0){/*do nothing*/}
if (images.size()<1) {
CV_Error(Error::StsBadArg, "Training data is not provided. Consider to add using addTrainingSample() function!");
}
if(strcmp(params.model_filename.c_str(),"")==0 && params.save_model){
CV_Error(Error::StsBadArg, "The model_filename parameter should be set!");
}
std::vector<std::vector<Point2f> > normalized;
Mat erode_kernel = getStructuringElement(MORPH_RECT, Size(3,3), Point(1,1));
Mat image;
int param_max_m = params.max_m;//550;
int param_max_n = params.max_n;//136;
AAM.scales = params.scales;
AAM.textures.resize(AAM.scales.size());
/*-------------- A. Load the training data---------*/
procrustesAnalysis(facePoints, normalized,AAM.s0);
/*-------------- B. Create the shape model---------*/
Mat s0_lin = linearize(AAM.s0).t() ;
// linearize all shapes data, all x and then all y for each shape
Mat M;
for(unsigned i=0;i<normalized.size();i++){
M.push_back(linearize(normalized[i]).t()-s0_lin);
}
/* get PCA Projection vectors */
Mat S;
getProjection(M.t(),S,param_max_n);
/* Create similarity eig*/
Mat shape_S,shape_Q;
calcSimilarityEig(AAM.s0,S,AAM.Q,AAM.S);
/* ----------C. Create the coordinate frame ------------*/
delaunay(AAM.s0,AAM.triangles);
for(size_t scale=0; scale<AAM.scales.size();scale++){
AAM.textures[scale].max_m = params.texture_max_m;//145;
if(params.verbose) printf("Training for scale %f ...\n", AAM.scales[scale]);
Mat s0_scaled_m = Mat(AAM.s0)/AAM.scales[scale]; // scale the shape
std::vector<Point2f> s0_scaled = s0_scaled_m.reshape(2); //convert to points
/*get the min and max of x and y coordinate*/
double min_x, max_x, min_y, max_y;
s0_scaled_m = s0_scaled_m.reshape(1);
Mat s0_scaled_x = s0_scaled_m.col(0);
Mat s0_scaled_y = s0_scaled_m.col(1);
minMaxIdx(s0_scaled_x, &min_x, &max_x);
minMaxIdx(s0_scaled_y, &min_y, &max_y);
std::vector<Point2f> base_shape = Mat(Mat(s0_scaled)-Scalar(min_x-2.0,min_y-2.0)).reshape(2);
AAM.textures[scale].base_shape = base_shape;
AAM.textures[scale].resolution = Rect(0,0,(int)ceil(max_x-min_x+3),(int)ceil(max_y-min_y+3));
Mat base_texture = createTextureBase(base_shape, AAM.triangles, AAM.textures[scale].resolution, AAM.textures[scale].textureIdx);
Mat mask1 = base_texture>0;
Mat mask2;
erode(mask1, mask1, erode_kernel);
erode(mask1, mask2, erode_kernel);
Mat warped;
std::vector<int> fe_map;
createMaskMapping(mask1,mask2, AAM.textures[scale].ind1, AAM.textures[scale].ind2,fe_map);//ok
/* ------------ Part D. Get textures -------------*/
Mat texture_feats, feat;
if(params.verbose) printf("(1/4) Feature extraction ...\n");
for(size_t i=0; i<images.size();i++){
if(params.verbose) printf("extract features from image #%i/%i\n", (int)(i+1), (int)images.size());
warped = warpImage(images[i],base_shape, facePoints[i], AAM.triangles, AAM.textures[scale].resolution,AAM.textures[scale].textureIdx);
feat = getFeature<uchar>(warped, AAM.textures[scale].ind1);
texture_feats.push_back(feat.t());
}
Mat T= texture_feats.t();
/* -------------- E. Create the texture model -----------------*/
reduce(T,AAM.textures[scale].A0,1, REDUCE_AVG);
if(params.verbose) printf("(2/4) Compute the feature average ...\n");
Mat A0_mtx = repeat(AAM.textures[scale].A0,1,T.cols);
Mat textures_normalized = T - A0_mtx;
if(params.verbose) printf("(3/4) Projecting the features ...\n");
getProjection(textures_normalized, AAM.textures[scale].A ,param_max_m);
AAM.textures[scale].AA0 = getFeature<float>(AAM.textures[scale].A0, fe_map);
if(params.verbose) printf("(4/4) Extraction of the eroded face features ...\n");
Mat U_data, ud;
for(int i =0;i<AAM.textures[scale].A.cols;i++){
Mat c = AAM.textures[scale].A.col(i);
ud = getFeature<float>(c,fe_map);
U_data.push_back(ud.t());
}
Mat U = U_data.t();
AAM.textures[scale].AA = orthonormal(U);
} // scale
images.clear();
if(params.save_model){
if(params.verbose) printf("Saving the model\n");
saveModel(params.model_filename);
}
isModelTrained = true;
if(params.verbose) printf("Training is completed\n");
}
bool FacemarkAAMImpl::fit( InputArray image, const std::vector<Rect>& roi, CV_OUT std::vector<std::vector<Point2f> >& _landmarks )
{
std::vector<Config> config; // empty
return fitConfig(image, roi, _landmarks, config);
}
bool FacemarkAAMImpl::fitConfig( InputArray image, const std::vector<Rect>& roi, std::vector<std::vector<Point2f> >& _landmarks, const std::vector<Config> &configs )
{
const std::vector<Rect> & faces = roi;
if(faces.size()<1) return false;
std::vector<std::vector<Point2f> > & landmarks = _landmarks;
landmarks.resize(faces.size());
Mat img = image.getMat();
if (! configs.empty()){
if (configs.size()!=faces.size()) {
CV_Error(Error::StsBadArg, "Number of faces and extra_parameters are different!");
}
for(size_t i=0; i<configs.size();i++){
fitImpl(img, landmarks[i], configs[i].R,configs[i].t, configs[i].scale, configs[i].model_scale_idx);
}
}else{
Mat R = Mat::eye(2, 2, CV_32F);
Point2f t = Point2f((float)(img.cols/2.0),(float)(img.rows/2.0));
float scale = 1.0;
for(unsigned i=0; i<faces.size();i++){
fitImpl(img, landmarks[i], R,t, scale);
}
}
return true;
}
bool FacemarkAAMImpl::fitImpl( const Mat image, std::vector<Point2f>& landmarks, const Mat R, const Point2f T, const float scale, int _scl){
if (landmarks.size()>0)
landmarks.clear();
CV_Assert(isModelTrained);
int param_n = params.n, param_m = params.m;
int scl = _scl<(int)AAM.scales.size()?_scl:(int)AAM.scales.size();
/*variables*/
std::vector<Point2f> s0 = Mat(Mat(AAM.s0)/AAM.scales[scl]).reshape(2);
/*pre-computation*/
Mat S = Mat(AAM.S, Range::all(), Range(0,param_n>AAM.S.cols?AAM.S.cols:param_n)).clone(); // chop the shape data
std::vector<std::vector<int> > Tp;
Mat Wx_dp, Wy_dp;
createWarpJacobian(S, AAM.Q, AAM.triangles, AAM.textures[scl],Wx_dp, Wy_dp, Tp);
std::vector<Point2f> s0_init = Mat(Mat(R*scale*AAM.scales[scl]*Mat(Mat(s0).reshape(1)).t()).t()).reshape(2);
std::vector<Point2f> curr_shape = Mat(Mat(s0_init)+Scalar(T.x,T.y));
curr_shape = Mat(1.0/scale*Mat(curr_shape)).reshape(2);
Mat imgray;
Mat img;
if(image.channels()>1){
cvtColor(image,imgray,COLOR_BGR2GRAY);
}else{
imgray = image;
}
resize(imgray,img,Size(int(image.cols/scale),int(image.rows/scale)), 0, 0, INTER_LINEAR_EXACT);// matlab use bicubic interpolation, the result is float numbers
/*chop the textures model*/
int maxCol = param_m;
if(AAM.textures[scl].A.cols<param_m)maxCol = AAM.textures[scl].A.cols;
if(AAM.textures[scl].AA.cols<maxCol)maxCol = AAM.textures[scl].AA.cols;
Mat A = Mat(AAM.textures[scl].A,Range(0,AAM.textures[scl].A.rows), Range(0,maxCol)).clone();
Mat AA = Mat(AAM.textures[scl].AA,Range(0,AAM.textures[scl].AA.rows), Range(0,maxCol)).clone();
/*iteratively update the fitting*/
Mat I, II, warped, c, gx, gy, Irec, Irec_feat, dc;
Mat refI, refII, refWarped, ref_c, ref_gx, ref_gy, refIrec, refIrec_feat, ref_dc ;
for(int t=0;t<params.n_iter;t++){
warped = warpImage(img,AAM.textures[scl].base_shape, curr_shape,
AAM.triangles,
AAM.textures[scl].resolution ,
AAM.textures[scl].textureIdx);
I = getFeature<uchar>(warped, AAM.textures[scl].ind1);
II = getFeature<uchar>(warped, AAM.textures[scl].ind2);
if(t==0){
c = A.t()*(I-AAM.textures[scl].A0); //little bit different to matlab, probably due to datatype
}else{
c = c+dc;
}
Irec_feat = (AAM.textures[scl].A0+A*c);
Irec = Mat::zeros(AAM.textures[scl].resolution.width, AAM.textures[scl].resolution.height, CV_32FC1);
for(int j=0;j<(int)AAM.textures[scl].ind1.size();j++){
Irec.at<float>(AAM.textures[scl].ind1[j]) = Irec_feat.at<float>(j);
}
Mat irec = Irec.t();
gradient(irec, gx, gy);
Mat Jc;
image_jacobian(Mat(gx.t()).reshape(0,1).t(),Mat(gy.t()).reshape(0,1).t(),Wx_dp, Wy_dp,Jc);
Mat J;
std::vector<float> Irec_vec;
for(size_t j=0;j<AAM.textures[scl].ind2.size();j++){
J.push_back(Jc.row(AAM.textures[scl].ind2[j]));
Irec_vec.push_back(Irec.at<float>(AAM.textures[scl].ind2[j]));
}
/*compute Jfsic and Hfsic*/
Mat Jfsic = J - AA*(AA.t()*J);
Mat Hfsic = Jfsic.t()*Jfsic;
Mat iHfsic;
invert(Hfsic, iHfsic);
/*compute dp dq and dc*/
Mat dqp = iHfsic*Jfsic.t()*(II-AAM.textures[scl].AA0);
dc = AA.t()*(II-Mat(Irec_vec)-J*dqp);
warpUpdate(curr_shape, dqp, s0,S, AAM.Q, AAM.triangles,Tp);
}
landmarks = Mat(scale*Mat(curr_shape)).reshape(2);
return true;
}
void FacemarkAAMImpl::saveModel(String s){
FileStorage fs(s.c_str(),FileStorage::WRITE_BASE64);
fs << "AAM_tri" << AAM.triangles;
fs << "scales" << AAM.scales;
fs << "s0" << AAM.s0;
fs << "S" << AAM.S;
fs << "Q" << AAM.Q;
String x;
for(int i=0;i< (int)AAM.scales.size();i++){
x = cv::format("scale%i_max_m",i);
fs << x << AAM.textures[i].max_m;
x = cv::format("scale%i_resolution",i);
fs << x << AAM.textures[i].resolution;
x = cv::format("scale%i_textureIdx",i);
fs << x << AAM.textures[i].textureIdx;
x = cv::format("scale%i_base_shape",i);
fs << x << AAM.textures[i].base_shape;
x = cv::format("scale%i_A",i);
fs << x << AAM.textures[i].A;
x = cv::format("scale%i_A0",i);
fs << x << AAM.textures[i].A0;
x = cv::format("scale%i_AA",i);
fs << x << AAM.textures[i].AA;
x = cv::format("scale%i_AA0",i);
fs << x << AAM.textures[i].AA0;
x = cv::format("scale%i_ind1",i);
fs << x << AAM.textures[i].ind1;
x = cv::format("scale%i_ind2",i);
fs << x << AAM.textures[i].ind2;
}
fs.release();
if(params.verbose) printf("The model is successfully saved! \n");
}
void FacemarkAAMImpl::loadModel(String s){
FileStorage fs(s.c_str(),FileStorage::READ);
String x;
fs["AAM_tri"] >> AAM.triangles;
fs["scales"] >> AAM.scales;
fs["s0"] >> AAM.s0;
fs["S"] >> AAM.S;
fs["Q"] >> AAM.Q;
AAM.textures.resize(AAM.scales.size());
for(int i=0;i< (int)AAM.scales.size();i++){
x = cv::format("scale%i_max_m",i);
fs[x] >> AAM.textures[i].max_m;
x = cv::format("scale%i_resolution",i);
fs[x] >> AAM.textures[i].resolution;
x = cv::format("scale%i_textureIdx",i);
fs[x] >> AAM.textures[i].textureIdx;
x = cv::format("scale%i_base_shape",i);
fs[x] >> AAM.textures[i].base_shape;
x = cv::format("scale%i_A",i);
fs[x] >> AAM.textures[i].A;
x = cv::format("scale%i_A0",i);
fs[x] >> AAM.textures[i].A0;
x = cv::format("scale%i_AA",i);
fs[x] >> AAM.textures[i].AA;
x = cv::format("scale%i_AA0",i);
fs[x] >> AAM.textures[i].AA0;
x = cv::format("scale%i_ind1",i);
fs[x] >> AAM.textures[i].ind1;
x = cv::format("scale%i_ind2",i);
fs[x] >> AAM.textures[i].ind2;
}
fs.release();
isModelTrained = true;
if(params.verbose) printf("the model has been loaded\n");
}
Mat FacemarkAAMImpl::procrustes(std::vector<Point2f> P, std::vector<Point2f> Q, Mat & rot, Scalar & trans, float & scale){
// calculate average
Scalar mx = mean(P);
Scalar my = mean(Q);
// zero centered data
Mat X0 = Mat(P) - mx;
Mat Y0 = Mat(Q) - my;
// calculate magnitude
Mat Xs, Ys;
multiply(X0,X0,Xs);
multiply(Y0,Y0,Ys);
// calculate the sum
Mat sumXs, sumYs;
reduce(Xs,sumXs, 0, REDUCE_SUM);
reduce(Ys,sumYs, 0, REDUCE_SUM);
//calculate the normrnd
double normX = sqrt(Mat(sumXs.reshape(1)).at<float>(0)+Mat(sumXs.reshape(1)).at<float>(1));
double normY = sqrt(Mat(sumYs.reshape(1)).at<float>(0)+Mat(sumYs.reshape(1)).at<float>(1));
//normalization
X0 = X0/normX;
Y0 = Y0/normY;
//reshape, convert to 2D Matrix
Mat Xn=X0.reshape(1);
Mat Yn=Y0.reshape(1);
//calculate the covariance matrix
Mat M = Xn.t()*Yn;
// decompose
Mat U,S,Vt;
SVD::compute(M, S, U, Vt);
// extract the transformations
scale = (S.at<float>(0)+S.at<float>(1))*(float)normX/(float)normY;
rot = Vt.t()*U.t();
Mat muX(mx),mX; muX.pop_back();muX.pop_back();
Mat muY(my),mY; muY.pop_back();muY.pop_back();
muX.convertTo(mX,CV_32FC1);
muY.convertTo(mY,CV_32FC1);
Mat t = mX.t()-scale*mY.t()*rot;
trans[0] = t.at<float>(0);
trans[1] = t.at<float>(1);
// calculate the recovered form
Mat Qmat = Mat(Q).reshape(1);
return Mat(scale*Qmat*rot+trans).clone();
}
void FacemarkAAMImpl::procrustesAnalysis(std::vector<std::vector<Point2f> > shapes, std::vector<std::vector<Point2f> > & normalized, std::vector<Point2f> & new_mean){
std::vector<Scalar> mean_every_shape;
mean_every_shape.resize(shapes.size());
Point2f temp;
// calculate the mean of every shape
for(size_t i=0; i< shapes.size();i++){
mean_every_shape[i] = mean(shapes[i]);
}
//normalize every shapes
Mat tShape;
normalized.clear();
for(size_t i=0; i< shapes.size();i++){
normalized.push_back((Mat)(Mat(shapes[i]) - mean_every_shape[i]));
}
// calculate the mean shape
std::vector<Point2f> mean_shape;
calcMeanShape(normalized, mean_shape);
// update the mean shape and normalized shapes iteratively
int maxIter = 100;
Mat R;
Scalar t;
float s;
Mat aligned;
for(int i=0;i<maxIter;i++){
// align
for(unsigned k=0;k< normalized.size();k++){
aligned=procrustes(mean_shape, normalized[k], R, t, s);
aligned.reshape(2).copyTo(normalized[k]);
}
//calc new mean
calcMeanShape(normalized, new_mean);
// align the new mean
aligned=procrustes(mean_shape, new_mean, R, t, s);
// update
aligned.reshape(2).copyTo(mean_shape);
}
}
void FacemarkAAMImpl::calcMeanShape(std::vector<std::vector<Point2f> > shapes,std::vector<Point2f> & mean){
mean.resize(shapes[0].size());
Point2f tmp;
for(unsigned i=0;i<shapes[0].size();i++){
tmp.x=0;
tmp.y=0;
for(unsigned k=0;k< shapes.size();k++){
tmp.x+= shapes[k][i].x;
tmp.y+= shapes[k][i].y;
}
tmp.x/=shapes.size();
tmp.y/=shapes.size();
mean[i] = tmp;
}
}
void FacemarkAAMImpl::getProjection(const Mat M, Mat & P, int n){
Mat U,S,Vt,S1, Ut;
int k;
if(M.rows < M.cols){
// SVD::compute(M*M.t(), S, U, Vt);
eigen(M*M.t(), S, Ut); U=Ut.t();
// find the minimum between number of non-zero eigval,
// compressed dim, row, and column
// threshold(S,S1,0.00001,1,THRESH_BINARY);
k= S.rows; //countNonZero(S1);
if(k>n)k=n;
if(k>M.rows)k=M.rows;
if(k>M.cols)k=M.cols;
// cut the column of eigen vector
U.colRange(0,k).copyTo(P);
}else{
// SVD::compute(M.t()*M, S, U, Vt);
eigen(M.t()*M, S, Ut);U=Ut.t();
// threshold(S,S1,0.00001,1,THRESH_BINARY);
k= S.rows; //countNonZero(S1);
if(k>n)k=n;
if(k>M.rows)k=M.rows;
if(k>M.cols)k=M.cols;
// cut the eigen values to k-amount
Mat D = Mat::zeros(k,k,CV_32FC1);
Mat diag = D.diag();
Mat s; pow(S,-0.5,s);
s(Range(0,k), Range::all()).copyTo(diag);
// cut the eigen vector to k-column,
P = Mat(M*U.colRange(0,k)*D).clone();
}
}
Mat FacemarkAAMImpl::orthonormal(Mat Mo){
Mat M;
Mo.convertTo(M,CV_32FC1);
// TODO: float precission is only 1e-7, but MATLAB version use thresh=2.2204e-16
float thresh = (float)2.2204e-6;
Mat O = Mat::zeros(M.rows, M.cols, CV_32FC1);
int k = 0; //storing index
Mat w,nv;
float n;
for(int i=0;i<M.cols;i++){
Mat v = M.col(i); // processed column to orthogonalize
// subtract projection over previous vectors
for(int j=0;j<k;j++){
Mat o=O.col(j);
w = v-o*(o.t()*v);
w.copyTo(v);
}
// only keep non zero vector
n = (float)norm(v);
if(n>thresh){
Mat ok=O.col(k);
// nv=v/n;
normalize(v,nv);
nv.copyTo(ok);
k+=1;
}
}
return O.colRange(0,k).clone();
}
void FacemarkAAMImpl::calcSimilarityEig(std::vector<Point2f> s0,Mat S, Mat & Q_orth, Mat & S_orth){
int npts = (int)s0.size();
Mat Q = Mat::zeros(2*npts,4,CV_32FC1);
Mat c0 = Q.col(0);
Mat c1 = Q.col(1);
Mat c2 = Q.col(2);
Mat c3 = Q.col(3);
/*c0 = s0(:)*/
Mat w = linearize(s0);
// w.convertTo(w, CV_64FC1);
w.copyTo(c0);
/*c1 = [-s0(npts:2*npts); s0(0:npts-1)]*/
Mat s0_mat = Mat(s0).reshape(1);
// s0_mat.convertTo(s0_mat, CV_64FC1);
Mat swapper = Mat::zeros(2,npts,CV_32FC1);
Mat s00 = s0_mat.col(0);
Mat s01 = s0_mat.col(1);
Mat sw0 = swapper.row(0);
Mat sw1 = swapper.row(1);
Mat(s00.t()).copyTo(sw1);
s01 = -s01;
Mat(s01.t()).copyTo(sw0);
Mat(swapper.reshape(1,2*npts)).copyTo(c1);
/*c2 - [ones(npts); zeros(npts)]*/
Mat ones = Mat::ones(1,npts,CV_32FC1);
Mat c2_mat = Mat::zeros(2,npts,CV_32FC1);
Mat c20 = c2_mat.row(0);
ones.copyTo(c20);
Mat(c2_mat.reshape(1,2*npts)).copyTo(c2);
/*c3 - [zeros(npts); ones(npts)]*/
Mat c3_mat = Mat::zeros(2,npts,CV_32FC1);
Mat c31 = c3_mat.row(1);
ones.copyTo(c31);
Mat(c3_mat.reshape(1,2*npts)).copyTo(c3);
Mat Qo = orthonormal(Q);
Mat all = Qo.t();
all.push_back(S.t());
Mat allOrth = orthonormal(all.t());
Q_orth = allOrth.colRange(0,4).clone();
S_orth = allOrth.colRange(4,allOrth.cols).clone();
}
inline Mat FacemarkAAMImpl::linearize(Mat s){ // all x values and then all y values
return Mat(s.reshape(1).t()).reshape(1,2*s.rows);
}
inline Mat FacemarkAAMImpl::linearize(std::vector<Point2f> s){ // all x values and then all y values
return linearize(Mat(s));
}
void FacemarkAAMImpl::delaunay(std::vector<Point2f> s, std::vector<Vec3i> & triangles)
{
triangles.clear();
std::vector<Vec6f> tp;
double min_x, max_x, min_y, max_y;
Mat S = Mat(s).reshape(1);
Mat s_x = S.col(0);
Mat s_y = S.col(1);
minMaxIdx(s_x, &min_x, &max_x);
minMaxIdx(s_y, &min_y, &max_y);
// TODO FIXIT Some triangles are lost
//Subdiv2D subdiv(Rect(cvFloor(min_x), cvFloor(min_y), cvCeil(max_x) - cvFloor(min_x), cvCeil(max_y) - cvFloor(min_y)));
Subdiv2D subdiv(Rect(cvFloor(min_x) - 10, cvFloor(min_y) - 10, cvCeil(max_x) - cvFloor(min_x) + 20, cvCeil(max_y) - cvFloor(min_y) + 20));
// map subdiv_verter -> original point (or the first alias)
std::vector<int> idx(s.size() + 4);
for (size_t i = 0; i < s.size(); ++i)
{
int vertex = subdiv.insert(s[i]);
if (idx.size() <= (size_t)vertex)
idx.resize(vertex + 1);
idx[vertex] = (int)i;
}
subdiv.getTriangleList(tp);
for (size_t i = 0; i < tp.size(); i++)
{
const Vec6f& t = tp[i];
//accept only vertex point
CV_Assert(
t[0]>=min_x && t[0]<=max_x && t[1]>=min_y && t[1]<=max_y &&
t[2]>=min_x && t[2]<=max_x && t[3]>=min_y && t[3]<=max_y &&
t[4]>=min_x && t[4]<=max_x && t[5]>=min_y && t[5]<=max_y
);
int tmp = 0, v1 = 0, v2 = 0, v3 = 0;
subdiv.locate(Point2f(t[0], t[1]), tmp, v1);
subdiv.locate(Point2f(t[2], t[3]), tmp, v2);
subdiv.locate(Point2f(t[4], t[5]), tmp, v3);
triangles.push_back(Vec3i(idx[v1], idx[v2], idx[v3]));
} // for
}
Mat FacemarkAAMImpl::createMask(std::vector<Point2f> base_shape, Rect res){
Mat mask = Mat::zeros(res.height, res.width, CV_8U);
std::vector<Point> hull;
std::vector<Point> shape;
Mat(base_shape).convertTo(shape, CV_32S);
convexHull(shape,hull);
fillConvexPoly(mask, &hull[0], (int)hull.size(), 255, 8 ,0);
return mask.clone();
}
Mat FacemarkAAMImpl::createTextureBase(std::vector<Point2f> shape, std::vector<Vec3i> triangles, Rect res, std::vector<std::vector<Point> > & textureIdx){
// max supported amount of triangles only 255
Mat mask = Mat::zeros(res.height, res.width, CV_8U);
std::vector<Point2f> p(3);
textureIdx.clear();
for(size_t i=0;i<triangles.size();i++){
p[0] = shape[triangles[i][0]];
p[1] = shape[triangles[i][1]];
p[2] = shape[triangles[i][2]];
std::vector<Point> polygon;
approxPolyDP(p,polygon, 1.0, true);
fillConvexPoly(mask, &polygon[0], (int)polygon.size(), (double)i+1,8,0 );
std::vector<Point> list;
for(int y=0;y<res.height;y++){
for(int x=0;x<res.width;x++){
if(mask.at<uchar>(y,x)==(uchar)(i+1)){
list.push_back(Point(x,y));
}
}
}
textureIdx.push_back(list);
}
return mask.clone();
}
Mat FacemarkAAMImpl::warpImage(
const Mat img, const std::vector<Point2f> target_shape,
const std::vector<Point2f> curr_shape, const std::vector<Vec3i> triangles,
const Rect res, const std::vector<std::vector<Point> > textureIdx)
{
// TODO: this part can be optimized, collect tranformation pair form all triangles first, then do one time remapping
Mat warped = Mat::zeros(res.height, res.width, CV_8U);
Mat warped2 = Mat::zeros(res.height, res.width, CV_8U);
Mat image,part, warped_part;
if(img.channels()>1){
cvtColor(img,image,COLOR_BGR2GRAY);
}else{
image = img;
}
Mat A,R,t;
A = Mat::zeros(2,3,CV_64F);
std::vector<Point2f> target(3),source(3);
std::vector<Point> polygon;
for(size_t i=0;i<triangles.size();i++){
target[0] = target_shape[triangles[i][0]];
target[1] = target_shape[triangles[i][1]];
target[2] = target_shape[triangles[i][2]];
source[0] = curr_shape[triangles[i][0]];
source[1] = curr_shape[triangles[i][1]];
source[2] = curr_shape[triangles[i][2]];
Mat target_mtx = Mat(target).reshape(1)-1.0;
Mat source_mtx = Mat(source).reshape(1)-1.0;
Mat U = target_mtx.col(0);
Mat V = target_mtx.col(1);
Mat X = source_mtx.col(0);
Mat Y = source_mtx.col(1);
double denominator = (target[1].x-target[0].x)*(target[2].y-target[0].y)-
(target[1].y-target[0].y)*(target[2].x-target[0].x);
// denominator = 1.0/denominator;
A.at<double>(0) = ((target[2].y-target[0].y)*(source[1].x-source[0].x)-
(target[1].y-target[0].y)*(source[2].x-source[0].x))/denominator;
A.at<double>(1) = ((target[1].x-target[0].x)*(source[2].x-source[0].x)-
(target[2].x-target[0].x)*(source[1].x-source[0].x))/denominator;
A.at<double>(2) =X.at<float>(0) + ((V.at<float>(0) * (U.at<float>(2) - U.at<float>(0)) - U.at<float>(0)*(V.at<float>(2) - V.at<float>(0))) * (X.at<float>(1) - X.at<float>(0)) + (U.at<float>(0) * (V.at<float>(1) - V.at<float>(0)) - V.at<float>(0)*(U.at<float>(1) - U.at<float>(0))) * (X.at<float>(2) - X.at<float>(0))) / denominator;
A.at<double>(3) =((V.at<float>(2) - V.at<float>(0)) * (Y.at<float>(1) - Y.at<float>(0)) - (V.at<float>(1) - V.at<float>(0)) * (Y.at<float>(2) - Y.at<float>(0))) / denominator;
A.at<double>(4) = ((U.at<float>(1) - U.at<float>(0)) * (Y.at<float>(2) - Y.at<float>(0)) - (U.at<float>(2) - U.at<float>(0)) * (Y.at<float>(1) - Y.at<float>(0))) / denominator;
A.at<double>(5) = Y.at<float>(0) + ((V.at<float>(0) * (U.at<float>(2) - U.at<float>(0)) - U.at<float>(0) * (V.at<float>(2) - V.at<float>(0))) * (Y.at<float>(1) - Y.at<float>(0)) + (U.at<float>(0) * (V.at<float>(1) - V.at<float>(0)) - V.at<float>(0)*(U.at<float>(1) - U.at<float>(0))) * (Y.at<float>(2) - Y.at<float>(0))) / denominator;
// A = getAffineTransform(target,source);
R=A.colRange(0,2);
t=A.colRange(2,3);
Mat pts_ori = Mat(textureIdx[i]).reshape(1);
Mat pts = pts_ori.t(); //matlab
Mat bx = pts_ori.col(0);
Mat by = pts_ori.col(1);
Mat base_ind = (by-1)*res.width+bx;
Mat pts_f;
pts.convertTo(pts_f,CV_64FC1);
pts_f.push_back(Mat::ones(1,(int)textureIdx[i].size(),CV_64FC1));
Mat trans = (A*pts_f).t();
Mat T; trans.convertTo(T, CV_32S); // this rounding make the result a little bit different to matlab
Mat mx = T.col(0);
Mat my = T.col(1);
Mat ind = (my-1)*image.cols+mx;
int maxIdx = image.rows*image.cols;
int idx;
for(int k=0;k<ind.rows;k++){
idx=ind.at<int>(k);
if(idx>=0 && idx<maxIdx){
warped.at<uchar>(base_ind.at<int>(k)) = (uchar)(image.at<uchar>(idx));
}
}
warped.copyTo(warped2);
}
return warped2.clone();
}
template <class T>
Mat FacemarkAAMImpl::getFeature(const Mat m, std::vector<int> map){
std::vector<float> feat;
Mat M = m.t();//matlab
for(size_t i=0;i<map.size();i++){
feat.push_back((float)M.at<T>(map[i]));
}
return Mat(feat).clone();
}
void FacemarkAAMImpl::createMaskMapping(const Mat m1, const Mat m2, std::vector<int> & ind1, std::vector<int> & ind2, std::vector<int> & ind3){
int cnt = 0, idx=0;
ind1.clear();
ind2.clear();
ind3.clear();
Mat mask = m1.t();//matlab
Mat mask2 = m2.t();//matlab
for(int i=0;i<mask.rows;i++){
for(int j=0;j<mask.cols;j++){
if(mask.at<uchar>(i,j)>0){
if(mask2.at<uchar>(i,j)>0){
ind2.push_back(idx);
ind3.push_back(cnt);
}
ind1.push_back(idx);
cnt +=1;
}
idx+=1;
} // j
} // i
}
void FacemarkAAMImpl::image_jacobian(const Mat gx, const Mat gy, const Mat Jx, const Mat Jy, Mat & G){
Mat Gx = repeat(gx,1,Jx.cols);
Mat Gy = repeat(gy,1,Jx.cols);
Mat G1,G2;
multiply(Gx,Jx,G1);
multiply(Gy,Jy,G2);
G=G1+G2;
}
void FacemarkAAMImpl::warpUpdate(std::vector<Point2f> & shape, Mat delta, std::vector<Point2f> s0, Mat S, Mat Q, std::vector<Vec3i> triangles,std::vector<std::vector<int> > Tp){
std::vector<Point2f> new_shape;
int nSimEig = 4;
/*get dr, dp and compute ds0*/
Mat dr = -Mat(delta, Range(0,nSimEig));
Mat dp = -Mat(delta, Range(nSimEig, delta.rows));
Mat ds0 = S*dp + Q*dr;
Mat ds0_mat = Mat::zeros((int)s0.size(),2, CV_32FC1);
Mat c0 = ds0_mat.col(0);
Mat c1 = ds0_mat.col(1);
Mat(ds0, Range(0,(int)s0.size())).copyTo(c0);
Mat(ds0, Range((int)s0.size(),(int)s0.size()*2)).copyTo(c1);
Mat s_new = computeWarpParts(shape,s0,ds0_mat, triangles, Tp);
Mat diff =linearize(Mat(s_new - Mat(s0).reshape(1)));
Mat r = Q.t()*diff;
Mat p = S.t()*diff;
Mat s = linearize(s0) +S*p + Q*r;
Mat(Mat(s.t()).reshape(0,2).t()).reshape(2).copyTo(shape);
}
Mat FacemarkAAMImpl::computeWarpParts(std::vector<Point2f> curr_shape,std::vector<Point2f> s0, Mat ds0, std::vector<Vec3i> triangles,std::vector<std::vector<int> > Tp){
std::vector<Point2f> new_shape;
std::vector<Point2f> ds = ds0.reshape(2);
float mx,my;
Mat A;
std::vector<Point2f> target(3),source(3);
std::vector<double> p(3);
p[2] = 1;
for(size_t i=0;i<s0.size();i++){
p[0] = s0[i].x + ds[i].x;
p[1] = s0[i].y + ds[i].y;
std::vector<Point2f> v;
std::vector<float>vx, vy;
for(size_t j=0;j<Tp[i].size();j++){
int idx = Tp[i][j];
target[0] = s0[triangles[idx][0]];
target[1] = s0[triangles[idx][1]];
target[2] = s0[triangles[idx][2]];
source[0] = curr_shape[triangles[idx][0]];
source[1] = curr_shape[triangles[idx][1]];
source[2] = curr_shape[triangles[idx][2]];
A = getAffineTransform(target,source);
Mat(A*Mat(p)).reshape(2).copyTo(v);
vx.push_back(v[0].x);
vy.push_back(v[0].y);
}// j
/*find the median*/
size_t n = vx.size()/2;
nth_element(vx.begin(), vx.begin()+n, vx.end());
mx = vx[n];
nth_element(vy.begin(), vy.begin()+n, vy.end());
my = vy[n];
new_shape.push_back(Point2f(mx,my));
} // s0.size()
return Mat(new_shape).reshape(1).clone();
}
void FacemarkAAMImpl::gradient(const Mat M, Mat & gx, Mat & gy){
gx = Mat::zeros(M.size(),CV_32FC1);
gy = Mat::zeros(M.size(),CV_32FC1);
/*gx*/
for(int i=0;i<M.rows;i++){
for(int j=0;j<M.cols;j++){
if(j>0 && j<M.cols-1){
gx.at<float>(i,j) = ((float)0.5)*(M.at<float>(i,j+1)-M.at<float>(i,j-1));
}else if (j==0){
gx.at<float>(i,j) = M.at<float>(i,j+1)-M.at<float>(i,j);
}else{
gx.at<float>(i,j) = M.at<float>(i,j)-M.at<float>(i,j-1);
}
}
}
/*gy*/
for(int i=0;i<M.rows;i++){
for(int j=0;j<M.cols;j++){
if(i>0 && i<M.rows-1){
gy.at<float>(i,j) = ((float)0.5)*(M.at<float>(i+1,j)-M.at<float>(i-1,j));
}else if (i==0){
gy.at<float>(i,j) = M.at<float>(i+1,j)-M.at<float>(i,j);
}else{
gy.at<float>(i,j) = M.at<float>(i,j)-M.at<float>(i-1,j);
}
}
}
}
void FacemarkAAMImpl::createWarpJacobian(Mat S, Mat Q, std::vector<Vec3i> triangles, Model::Texture & T, Mat & Wx_dp, Mat & Wy_dp, std::vector<std::vector<int> > & Tp){
std::vector<Point2f> base_shape = T.base_shape;
Rect resolution = T.resolution;
std::vector<std::vector<int> >triangles_on_a_point;
int npts = (int)base_shape.size();
Mat dW_dxdyt ;
/*get triangles for each point*/
std::vector<int> trianglesIdx;
triangles_on_a_point.resize(npts);
for(int i=0;i<(int)triangles.size();i++){
triangles_on_a_point[triangles[i][0]].push_back(i);
triangles_on_a_point[triangles[i][1]].push_back(i);
triangles_on_a_point[triangles[i][2]].push_back(i);
}
Tp = triangles_on_a_point;
/*calculate dW_dxdy*/
float v0x,v0y,v1x,v1y,v2x,v2y, denominator;
for(int k=0;k<npts;k++){
Mat acc = Mat::zeros(resolution.height, resolution.width, CV_32F);
/*for each triangle on k-th point*/
for(size_t i=0;i<triangles_on_a_point[k].size();i++){
int tId = triangles_on_a_point[k][i];
Vec3i v;
if(triangles[tId][0]==k ){
v=Vec3i(triangles[tId][0],triangles[tId][1],triangles[tId][2]);
}else if(triangles[tId][1]==k){
v=Vec3i(triangles[tId][1],triangles[tId][0],triangles[tId][2]);
}else{
v=Vec3i(triangles[tId][2],triangles[tId][0],triangles[tId][1]);
}
v0x = base_shape[v[0]].x;
v0y = base_shape[v[0]].y;
v1x = base_shape[v[1]].x;
v1y = base_shape[v[1]].y;
v2x = base_shape[v[2]].x;
v2y = base_shape[v[2]].y;
denominator = (v1x-v0x)*(v2y-v0y)-(v1y-v0y)*(v2x-v0x);
Mat pixels = Mat(T.textureIdx[tId]).reshape(1); // same, just different order
Mat p;
pixels.convertTo(p,CV_32F, 1.0,1.0); //matlab use offset
Mat x = p.col(0);
Mat y = p.col(1);
Mat alpha = (x-v0x)*(v2y-v0y)-(y-v0y)*(v2x-v0x);
Mat beta = (v1x-v0x)*(y-v0y)-(v1y-v0y)*(x-v0x);
Mat res = 1.0 - alpha/denominator - beta/denominator; // same just different order
/*remap to image form*/
Mat dx = Mat::zeros(resolution.height, resolution.width, CV_32F);
for(int j=0;j<res.rows;j++){
dx.at<float>((int)(y.at<float>(j)-1.0), (int)(x.at<float>(j)-1.0)) = res.at<float>(j); // matlab use offset
};
acc = acc+dx;
}
Mat vectorized = Mat(acc.t()).reshape(0,1);
dW_dxdyt.push_back(vectorized.clone());
}// k
Mat dx_dp;
hconcat(Q, S, dx_dp);
Mat dW_dxdy = dW_dxdyt.t();
Wx_dp = dW_dxdy* Mat(dx_dp,Range(0,npts));
Wy_dp = dW_dxdy* Mat(dx_dp,Range(npts,2*npts));
} //createWarpJacobian
} /* namespace face */
} /* namespace cv */