void ShapeModel::viewShapeModelUpdate(ModelViewInfo *pInfo)
{
Mat_< double > paramV;
paramV.create(this->nShapeParams, 1);
for (int i=0;i<nShapeParams;i++){
paramV(i, 0) = (pInfo->vList[i]/30.0 - 0.5)*6*
sqrt(pcaShape->eigenvalues.at<double>(i, 0));
}
Mat_<cv::Vec3b> img;
ModelImage s;
s.setShapeInfo(&shapeInfo);
s.loadTrainImage(Mat_<unsigned char>::ones(190*2, 160*2)*255);
projectParamToShape(paramV, s.shapeVec);
SimilarityTrans st = s.shapeVec.getShapeTransformFitingSize(
Size(320, 380));
s.buildFromShapeVec(st, 1000);
img = s.show(0, -1, false);
imshow("Viewing Shape Model", img);
}
void ASMModel::doSearch(const cv::Mat& img, FitResult& fitResult, int verbose)
{
// Step 2: Ensure it is a grayscale image
Mat grayImg;
if (img.channels() == 3){
cv::cvtColor(img, grayImg, CV_BGR2GRAY);
}
else
grayImg = img;
// Step 3: Resize each face image
Mat resizedImg;
// Resize the image to proper size
double ratio;
ratio = sqrt( double(40000) / (grayImg.rows * grayImg.cols));
cv::resize(grayImg, resizedImg, cv::Size(grayImg.cols*ratio, grayImg.rows*ratio));
puts("Start fitting...");
ModelImage curSearch;
curSearch.setShapeInfo( &shapeInfo );
curSearch.loadTrainImage(resizedImg);
fitResult.params = Mat_<double>::zeros(nShapeParams, 1);
ShapeVec &sv = curSearch.shapeVec;
ShapeVec shape_old;
projectParamToShape(fitResult.params, sv);
SimilarityTrans st = sv.getShapeTransformFitingSize(resizedImg.size(),
searchScaleRatio,
searchInitXOffset,
searchInitYOffset);
fitResult.transformation = st;
curSearch.buildFromShapeVec(st);
pyramidLevel = 2;
int k=localFeatureRad;
ns=4;
// sum of offsets of current iteration.
int totalOffset;
if (verbose >= ASM_FIT_VERBOSE_AT_LEVEL)
curSearch.show();
// Update each point
vector< Point_< int > > V;
for (int l=this->pyramidLevel; l>=0; l--){
if (verbose >= ASM_FIT_VERBOSE_AT_LEVEL)
printf("Level %d\n", l);
Mat_<double> img=curSearch.getDerivImage(l);
// printf("Image size: %dx%d\n", img.cols, img.rows);
// at most 5 iterations for each level
int runT;
double avgMov;
// [zl] this number has funky effects and also asjust speed a bit.
for (runT=0; runT<3; runT++){
// Backup current shape
shape_old.fromPointList(curSearch.points);
totalOffset = 0;
vector< Point_< int > > bestEP(nMarkPoints);
for (int i=0; i<this->nMarkPoints; i++){
if (verbose >= ASM_FIT_VERBOSE_AT_POINT)
printf("Dealing point %d...\n", i);
Mat_< double > nrmV(2*k+1, 1);
double curBest=-1, ct;
int bestI = 0;
double absSum;
for (int e=ns; e>=-ns; e--){
curSearch.getPointsOnNorm(i, k, l, V, 2*searchStepAreaRatio, e);
absSum = 0;
for (int j=-k;j<=k;j++){
nrmV(j+k, 0) = img(V[j+k]);
absSum += fabs(nrmV(j+k, 0));
}
nrmV *= 1/absSum;
ct = cv::Mahalanobis(nrmV, this->meanG[l][i], this->iCovarG[l][i]);
// printf("absSum: %lf, ct: %lf\n", absSum, ct);
if (verbose >= ASM_FIT_VERBOSE_AT_POINT)
curSearch.show(l, i, true, e);
if (ct<curBest || curBest<0){
curBest = ct;
bestI = e;
bestEP[i] = V[k];
}
}
// printf("best e: %d\n", bestI);
// bestEP[i] = V[bestI+(ns+k)];
totalOffset += abs(bestI);
if (verbose >= ASM_FIT_VERBOSE_AT_POINT)
curSearch.show(l, i, true, bestI);
}
for (int i=0;i<nMarkPoints;i++){
curSearch.points[i] = bestEP[i];
curSearch.points[i].x <<= l;
if (l>0) curSearch.points[i].x += (1<<(l-1));
curSearch.points[i].y <<= l;
if (l>0) curSearch.points[i].y += (1<<(l-1));
}
curSearch.shapeVec.fromPointList(curSearch.points);
if (verbose >= ASM_FIT_VERBOSE_AT_ITERATION)
curSearch.show(l);
// Project to PCA model and then back
//findParamForShape(curSearch.shapeVec, fitResult);
findParamForShapeBTSM(curSearch.shapeVec, shape_old, fitResult, fitResult, l);
pcaPyr[l].backProject(fitResult.params, sv);
// Reconstruct new shape
curSearch.buildFromShapeVec(fitResult.transformation);
avgMov = (double)totalOffset/nMarkPoints;
if (verbose >= ASM_FIT_VERBOSE_AT_ITERATION){
printf("Iter %d: Average offset: %.3f\n", runT+1, avgMov);
curSearch.show(l);
}
if (avgMov < 1.3){
runT++;
break;
}
}
if (verbose == ASM_FIT_VERBOSE_AT_LEVEL){
printf("%d iterations. average offset for last iter: %.3f\n", runT, avgMov);
curSearch.show(l);
}
}
SimilarityTrans s2;
s2.a = 1/ratio;
fitResult.transformation = s2 * fitResult.transformation;
}