void rotate_dataset(const command_line_parser& parser)
{
image_dataset_metadata::dataset metadata;
const string datasource = parser[0];
load_image_dataset_metadata(metadata,datasource);
double angle = get_option(parser, "rotate", 0);
// Set the current directory to be the one that contains the
// metadata file. We do this because the file might contain
// file paths which are relative to this folder.
set_current_dir(get_parent_directory(file(datasource)));
const string file_prefix = "rotated_"+ cast_to_string(angle) + "_";
const string metadata_filename = get_parent_directory(file(datasource)).full_name() +
directory::get_separator() + file_prefix + file(datasource).name();
array2d<rgb_pixel> img, temp;
for (unsigned long i = 0; i < metadata.images.size(); ++i)
{
file f(metadata.images[i].filename);
const string filename = get_parent_directory(f).full_name() + directory::get_separator() + file_prefix + to_png_name(f.name());
load_image(img, metadata.images[i].filename);
const point_transform_affine tran = rotate_image(img, temp, angle*pi/180);
save_png(temp, filename);
for (unsigned long j = 0; j < metadata.images[i].boxes.size(); ++j)
{
const rectangle rect = metadata.images[i].boxes[j].rect;
rectangle newrect;
newrect += tran(rect.tl_corner());
newrect += tran(rect.tr_corner());
newrect += tran(rect.bl_corner());
newrect += tran(rect.br_corner());
// now make newrect have the same area as the starting rect.
double ratio = std::sqrt(rect.area()/(double)newrect.area());
newrect = centered_rect(newrect, newrect.width()*ratio, newrect.height()*ratio);
metadata.images[i].boxes[j].rect = newrect;
// rotate all the object parts
std::map<std::string,point>::iterator k;
for (k = metadata.images[i].boxes[j].parts.begin(); k != metadata.images[i].boxes[j].parts.end(); ++k)
{
k->second = tran(k->second);
}
}
metadata.images[i].filename = filename;
}
save_image_dataset_metadata(metadata, metadata_filename);
}
unsigned long area(const rectangle& r) { return r.area(); }
int resample_dataset(const command_line_parser& parser)
{
if (parser.number_of_arguments() != 1)
{
cerr << "The --resample option requires you to give one XML file on the command line." << endl;
return EXIT_FAILURE;
}
const size_t obj_size = get_option(parser,"resample",100*100);
const double margin_scale = 2.5; // cropped image will be this times wider than the object.
const size_t image_size = obj_size*margin_scale*margin_scale;
dlib::image_dataset_metadata::dataset data, resampled_data;
resampled_data.comment = data.comment;
resampled_data.name = data.name + " RESAMPLED";
load_image_dataset_metadata(data, parser[0]);
locally_change_current_dir chdir(get_parent_directory(file(parser[0])));
console_progress_indicator pbar(data.images.size());
for (unsigned long i = 0; i < data.images.size(); ++i)
{
// don't even bother loading images that don't have objects.
if (data.images[i].boxes.size() == 0)
continue;
pbar.print_status(i);
array2d<rgb_pixel> img, chip;
load_image(img, data.images[i].filename);
// figure out what chips we want to take from this image
for (unsigned long j = 0; j < data.images[i].boxes.size(); ++j)
{
const rectangle rect = data.images[i].boxes[j].rect;
if (data.images[i].boxes[j].ignore || !get_rect(img).contains(rect))
continue;
const rectangle crop_rect = centered_rect(rect, rect.width()*margin_scale, rect.height()*margin_scale);
// skip crops that have a lot of border pixels
if (get_rect(img).intersect(crop_rect).area() < crop_rect.area()*0.8)
continue;
const rectangle_transform tform = get_mapping_to_chip(chip_details(crop_rect, image_size));
extract_image_chip(img, chip_details(crop_rect, image_size), chip);
image_dataset_metadata::image dimg;
// Now transform the boxes to the crop and also mark them as ignored if they
// have already been cropped out or are outside the crop.
for (size_t k = 0; k < data.images[i].boxes.size(); ++k)
{
image_dataset_metadata::box box = data.images[i].boxes[k];
// ignore boxes outside the cropped image
if (crop_rect.intersect(box.rect).area() == 0)
continue;
// mark boxes we include in the crop as ignored. Also mark boxes that
// aren't totally within the crop as ignored.
if (crop_rect.contains(grow_rect(box.rect,10)))
data.images[i].boxes[k].ignore = true;
else
box.ignore = true;
box.rect = tform(box.rect);
for (auto&& p : box.parts)
p.second = tform.get_tform()(p.second);
dimg.boxes.push_back(box);
}
dimg.filename = data.images[i].filename + "RESAMPLED"+cast_to_string(j)+".jpg";
save_jpeg(chip,dimg.filename, 98);
resampled_data.images.push_back(dimg);
}
}
save_image_dataset_metadata(resampled_data, parser[0] + ".RESAMPLED.xml");
return EXIT_SUCCESS;
}
int resample_dataset(const command_line_parser& parser)
{
if (parser.number_of_arguments() != 1)
{
cerr << "The --resample option requires you to give one XML file on the command line." << endl;
return EXIT_FAILURE;
}
const size_t obj_size = get_option(parser,"cropped-object-size",100*100);
const double margin_scale = get_option(parser,"crop-size",2.5); // cropped image will be this times wider than the object.
const unsigned long min_object_size = get_option(parser,"min-object-size",1);
const bool one_object_per_image = parser.option("one-object-per-image");
dlib::image_dataset_metadata::dataset data, resampled_data;
std::ostringstream sout;
sout << "\nThe --resample parameters which generated this dataset were:" << endl;
sout << " cropped-object-size: "<< obj_size << endl;
sout << " crop-size: "<< margin_scale << endl;
sout << " min-object-size: "<< min_object_size << endl;
if (one_object_per_image)
sout << " one_object_per_image: true" << endl;
resampled_data.comment = data.comment + sout.str();
resampled_data.name = data.name + " RESAMPLED";
load_image_dataset_metadata(data, parser[0]);
locally_change_current_dir chdir(get_parent_directory(file(parser[0])));
dlib::rand rnd;
const size_t image_size = std::round(std::sqrt(obj_size*margin_scale*margin_scale));
const chip_dims cdims(image_size, image_size);
console_progress_indicator pbar(data.images.size());
for (unsigned long i = 0; i < data.images.size(); ++i)
{
// don't even bother loading images that don't have objects.
if (data.images[i].boxes.size() == 0)
continue;
pbar.print_status(i);
array2d<rgb_pixel> img, chip;
load_image(img, data.images[i].filename);
// figure out what chips we want to take from this image
for (unsigned long j = 0; j < data.images[i].boxes.size(); ++j)
{
const rectangle rect = data.images[i].boxes[j].rect;
if (data.images[i].boxes[j].ignore || rect.area() < min_object_size)
continue;
const auto max_dim = std::max(rect.width(), rect.height());
const double rand_scale_perturb = 1 - 0.3*(rnd.get_random_double()-0.5);
const rectangle crop_rect = centered_rect(rect, max_dim*margin_scale*rand_scale_perturb, max_dim*margin_scale*rand_scale_perturb);
const rectangle_transform tform = get_mapping_to_chip(chip_details(crop_rect, cdims));
extract_image_chip(img, chip_details(crop_rect, cdims), chip);
image_dataset_metadata::image dimg;
// Now transform the boxes to the crop and also mark them as ignored if they
// have already been cropped out or are outside the crop.
for (size_t k = 0; k < data.images[i].boxes.size(); ++k)
{
image_dataset_metadata::box box = data.images[i].boxes[k];
// ignore boxes outside the cropped image
if (crop_rect.intersect(box.rect).area() == 0)
continue;
// mark boxes we include in the crop as ignored. Also mark boxes that
// aren't totally within the crop as ignored.
if (crop_rect.contains(grow_rect(box.rect,10)) && (!one_object_per_image || k==j))
data.images[i].boxes[k].ignore = true;
else
box.ignore = true;
if (box.rect.area() < min_object_size)
box.ignore = true;
box.rect = tform(box.rect);
for (auto&& p : box.parts)
p.second = tform.get_tform()(p.second);
dimg.boxes.push_back(box);
}
// Put a 64bit hash of the image data into the name to make sure there are no
// file name conflicts.
std::ostringstream sout;
sout << hex << murmur_hash3_128bit(&chip[0][0], chip.size()*sizeof(chip[0][0])).second;
dimg.filename = data.images[i].filename + "_RESAMPLED_"+sout.str()+".png";
if (parser.option("jpg"))
{
dimg.filename = to_jpg_name(dimg.filename);
save_jpeg(chip,dimg.filename, JPEG_QUALITY);
}
else
{
save_png(chip,dimg.filename);
}
resampled_data.images.push_back(dimg);
}
}
save_image_dataset_metadata(resampled_data, parser[0] + ".RESAMPLED.xml");
return EXIT_SUCCESS;
}
