PcaModel PcaModel::loadStatismoModel(path h5file, PcaModel::ModelType modelType)
{
logging::Logger logger = Loggers->getLogger("shapemodels");
PcaModel model;
// Load the shape or color model from the .h5 file
string h5GroupType;
if (modelType == ModelType::SHAPE) {
h5GroupType = "shape";
} else if (modelType == ModelType::COLOR) {
h5GroupType = "color";
}
H5::H5File h5Model;
try {
h5Model = H5::H5File(h5file.string(), H5F_ACC_RDONLY);
}
catch (H5::Exception& e) {
string errorMessage = "Could not open HDF5 file: " + string(e.getCDetailMsg());
logger.error(errorMessage);
throw errorMessage;
}
// Load either the shape or texture mean
string h5Group = "/" + h5GroupType + "/model";
H5::Group modelReconstructive = h5Model.openGroup(h5Group);
// Read the mean
H5::DataSet dsMean = modelReconstructive.openDataSet("./mean");
hsize_t dims[2];
dsMean.getSpace().getSimpleExtentDims(dims, NULL); // dsMean.getSpace() leaks memory... maybe a hdf5 bug, maybe vlenReclaim(...) could be a fix. No idea.
//H5::DataSpace dsp = dsMean.getSpace();
//dsp.close();
Loggers->getLogger("shapemodels").debug("Dimensions of the model mean: " + lexical_cast<string>(dims[0]));
model.mean = Mat(1, dims[0], CV_32FC1); // Use a row-vector, because of faster memory access and I'm not sure the memory block is allocated contiguously if we have multiple rows.
dsMean.read(model.mean.ptr<float>(0), H5::PredType::NATIVE_FLOAT);
model.mean = model.mean.t(); // Transpose it to a col-vector
dsMean.close();
// Read the eigenvalues
dsMean = modelReconstructive.openDataSet("./pcaVariance");
dsMean.getSpace().getSimpleExtentDims(dims, NULL);
Loggers->getLogger("shapemodels").debug("Dimensions of the pcaVariance: " + lexical_cast<string>(dims[0]));
model.eigenvalues = Mat(1, dims[0], CV_32FC1);
dsMean.read(model.eigenvalues.ptr<float>(0), H5::PredType::NATIVE_FLOAT);
model.eigenvalues = model.eigenvalues.t();
dsMean.close();
// Read the PCA basis matrix
dsMean = modelReconstructive.openDataSet("./pcaBasis");
dsMean.getSpace().getSimpleExtentDims(dims, NULL);
Loggers->getLogger("shapemodels").debug("Dimensions of the PCA basis matrix: " + lexical_cast<string>(dims[0]) + ", " + lexical_cast<string>(dims[1]));
model.pcaBasis = Mat(dims[0], dims[1], CV_32FC1);
dsMean.read(model.pcaBasis.ptr<float>(0), H5::PredType::NATIVE_FLOAT);
dsMean.close();
modelReconstructive.close(); // close the model-group
// Read the noise variance (not implemented)
/*dsMean = modelReconstructive.openDataSet("./noiseVariance");
float noiseVariance = 10.0f;
dsMean.read(&noiseVariance, H5::PredType::NATIVE_FLOAT);
dsMean.close(); */
// Read the triangle-list
string representerGroupName = "/" + h5GroupType + "/representer";
H5::Group representerGroup = h5Model.openGroup(representerGroupName);
dsMean = representerGroup.openDataSet("./reference-mesh/triangle-list");
dsMean.getSpace().getSimpleExtentDims(dims, NULL);
Loggers->getLogger("shapemodels").debug("Dimensions of the triangle-list: " + lexical_cast<string>(dims[0]) + ", " + lexical_cast<string>(dims[1]));
Mat triangles(dims[0], dims[1], CV_32SC1);
dsMean.read(triangles.ptr<int>(0), H5::PredType::NATIVE_INT32);
dsMean.close();
representerGroup.close();
model.triangleList.resize(triangles.rows);
for (unsigned int i = 0; i < model.triangleList.size(); ++i) {
model.triangleList[i][0] = triangles.at<int>(i, 0);
model.triangleList[i][1] = triangles.at<int>(i, 1);
model.triangleList[i][2] = triangles.at<int>(i, 2);
}
// Load the landmarks mappings:
// load the reference-mesh
representerGroup = h5Model.openGroup(representerGroupName);
dsMean = representerGroup.openDataSet("./reference-mesh/vertex-coordinates");
dsMean.getSpace().getSimpleExtentDims(dims, NULL);
Loggers->getLogger("shapemodels").debug("Dimensions of the reference-mesh vertex-coordinates matrix: " + lexical_cast<string>(dims[0]) + ", " + lexical_cast<string>(dims[1]));
Mat referenceMesh(dims[0], dims[1], CV_32FC1);
dsMean.read(referenceMesh.ptr<float>(0), H5::PredType::NATIVE_FLOAT);
dsMean.close();
representerGroup.close();
// convert to 3 vectors with the x, y and z coordinates for easy searching
vector<float> refx(referenceMesh.col(0).clone());
vector<float> refy(referenceMesh.col(1).clone());
vector<float> refz(referenceMesh.col(2).clone());
// load the landmarks info (mapping name <-> reference (x, y, z)-coords)
H5::Group landmarksGroup = h5Model.openGroup("/metadata/landmarks");
dsMean = landmarksGroup.openDataSet("./text");
H5std_string outputString;
Loggers->getLogger("shapemodels").debug("Reading landmark information from the model.");
dsMean.read(outputString, dsMean.getStrType());
dsMean.close();
landmarksGroup.close();
vector<string> landmarkLines;
boost::split(landmarkLines, outputString, boost::is_any_of("\n"), boost::token_compress_on);
for (const auto& l : landmarkLines) {
if (l == "") {
continue;
}
vector<string> line;
boost::split(line, l, boost::is_any_of(" "), boost::token_compress_on);
string name = line[0];
int visibility = lexical_cast<int>(line[1]);
float x = lexical_cast<float>(line[2]);
float y = lexical_cast<float>(line[3]);
float z = lexical_cast<float>(line[4]);
// Find the x, y and z values in the reference
const auto ivx = std::find(begin(refx), end(refx), x);
const auto ivy = std::find(begin(refy), end(refy), y);
const auto ivz = std::find(begin(refz), end(refz), z);
// TODO Check for .end()!
const auto vertexIdX = std::distance(begin(refx), ivx);
const auto vertexIdY = std::distance(begin(refy), ivy);
const auto vertexIdZ = std::distance(begin(refz), ivz);
// assert vx=vy=vz
// Hmm this is not perfect. If there's another vertex where 1 or 2 coords are the same, it fails.
// We should do the search differently: Find _all_ the vertices that are equal, then take the one that has the right x, y and z.
model.landmarkVertexMap.insert(make_pair(name, vertexIdX));
}
h5Model.close();
return model;
}
Matrix4D ReflectX()
{
Matrix4D refx(1, 0, 0, 0, 0, -1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 1);
return refx;
}
