void ImageData::SaveDescriptors(bool clear)
{
if (m_descriptors.empty()) return;
util::SaveContainerToFileBinary(m_filename + ".desc", m_descriptors);
if (clear) ClearDescriptors();
}
void plSDLMgr::read(hsStream* S) {
ClearDescriptors();
fDescriptors.resize(S->readShort());
for (size_t i=0; i<fDescriptors.size(); i++) {
fDescriptors[i] = new plStateDescriptor();
fDescriptors[i]->read(S);
}
// Propagate types on SDVars
for (size_t i=0; i<fDescriptors.size(); i++) {
plStateDescriptor* desc = fDescriptors[i];
for (size_t j=0; j<desc->getNumVars(); j++) {
plVarDescriptor* var = desc->get(j);
if (var->getType() == plVarDescriptor::kStateDescriptor)
var->setStateDesc(GetDescriptor(var->getStateDescType(),
var->getStateDescVer()));
}
}
}
void ImageData::DetectFeatures(const Options& opts)
{
// Cleanup data structures
ClearDescriptors();
m_keys.clear();
std::vector<cv::KeyPoint> keys;
const auto img = cv::imread(m_filename);
cv::Mat grey_img;
cv::cvtColor(img, grey_img, CV_BGR2GRAY);
switch (opts.feature_type)
{
case 0: // Detect YAPE/Daisy features
{
cv::YAPE YapeDet{opts.yape_radius,
opts.yape_threshold,
opts.yape_octaves,
opts.yape_views,
opts.yape_base_feature_size,
opts.yape_clustering_distance};
YapeDet(grey_img, keys);
daisy DaisyDesc{};
DaisyDesc.set_image(grey_img.data, m_height, m_width);
DaisyDesc.set_parameters(opts.daisy_radius,
opts.daisy_radius_quantization,
opts.daisy_angular_quantization,
opts.daisy_histogram_quantization);
DaisyDesc.verbose(0);
DaisyDesc.initialize_single_descriptor_mode();
m_desc_size = DaisyDesc.descriptor_size();
m_descriptors.reserve(m_keys.size() * m_desc_size);
// Compute Daisy descriptors at provided locations
for (const auto& key : keys)
{
std::vector<float> descriptor(m_desc_size);
DaisyDesc.get_descriptor(key.pt.y, key.pt.x, 35, descriptor.data());
m_descriptors.insert(m_descriptors.end(), descriptor.begin(), descriptor.end());
}
break;
}
case 1: // Detect SURF features
{
if (opts.surf_desc_extended) m_desc_size = 128;
else m_desc_size = 64;
cv::SURF Surf{opts.surf_det_hessian_threshold,
opts.surf_common_octaves,
opts.surf_common_octave_layers,
opts.surf_desc_extended,
opts.surf_desc_upright};
Surf(grey_img, cv::Mat{}, keys, m_descriptors);
break;
}
case 2: // Detect AKAZE features
{
m_desc_size = opts.akaze_descriptor_size;
cv::Mat working_img;
grey_img.convertTo(working_img, CV_32F, 1.0 / 255.0, 0);
AKAZEOptions options;
options.img_width = grey_img.cols;
options.img_height = grey_img.rows;
options.soffset = DEFAULT_SCALE_OFFSET;
options.omin = DEFAULT_OCTAVE_MIN;
options.omax = DEFAULT_OCTAVE_MAX;
options.nsublevels = DEFAULT_NSUBLEVELS;
options.dthreshold = opts.akaze_threshold;
options.diffusivity = DEFAULT_DIFFUSIVITY_TYPE;
options.descriptor = DEFAULT_DESCRIPTOR;
options.descriptor_size = opts.akaze_descriptor_size;
options.descriptor_channels = DEFAULT_LDB_CHANNELS;
options.descriptor_pattern_size = DEFAULT_LDB_PATTERN_SIZE;
options.sderivatives = DEFAULT_SIGMA_SMOOTHING_DERIVATIVES;
options.save_scale_space = DEFAULT_SAVE_SCALE_SPACE;
options.save_keypoints = DEFAULT_SAVE_KEYPOINTS;
options.verbosity = DEFAULT_VERBOSITY;
AKAZE akaze{options};
akaze.Create_Nonlinear_Scale_Space(working_img);
akaze.Feature_Detection(keys);
akaze.Compute_Descriptors(keys, m_descriptors_akaze);
break;
}
}
// Convert the keypoints from OpenCV's format to internal
m_keys.reserve(keys.size());
const float x_correction_factor = 0.5 * m_width;
const float y_correction_factor = 0.5 * m_height - 1.0;
for (const auto& key : keys)
{
const float x = key.pt.x - x_correction_factor;
const float y = y_correction_factor - key.pt.y;
const int xf = static_cast<int>(std::floor(key.pt.x)), yf = static_cast<int>(std::floor(key.pt.y));
const uchar* ptr = img.ptr<uchar>(yf);
m_keys.push_back(KeyPoint{x, y, ptr[3 * xf + 2], ptr[3 * xf + 1], ptr[3 * xf]});
}
SaveDescriptors(true);
}
