bool tracking_thread<Tnet>::get_data(vector<bbox*> &bboxes2,
idx<ubyte> &frame2, idx<ubyte> &tpl2) {
// lock data
pthread_mutex_lock(&mutex_out);
// only read data if it has been updated
if (!out_updated) {
// unlock data
pthread_mutex_unlock(&mutex_out);
return false;
}
// clear bboxes
for (ibox = bboxes2.begin(); ibox != bboxes2.end(); ++ibox) {
if (*ibox)
delete *ibox;
}
bboxes2.clear();
// copy bboxes pointers (now responsible for deleting them).
for (ibox = bboxes.begin(); ibox != bboxes.end(); ++ibox) {
bboxes2.push_back(*ibox);
}
// check frame is correctly allocated, if not, allocate.
if (frame2.order() != frame.order())
frame2 = idx<ubyte>(frame.get_idxdim());
else if (frame2.get_idxdim() != frame.get_idxdim())
frame2.resize(frame.get_idxdim());
// copy frame
idx_copy(frame, frame2);
// check tpl is correctly allocated, if not, allocate.
if (tpl2.order() != tpl.order())
tpl2 = idx<ubyte>(tpl.get_idxdim());
else if (tpl2.get_idxdim() != tpl.get_idxdim())
tpl2.resize(tpl.get_idxdim());
// copy tpl
idx_copy(tpl, tpl2);
// reset updated flag
out_updated = false;
// unlock data
pthread_mutex_unlock(&mutex_out);
// confirm that we copied data.
return true;
}
void padder<T>::pad(idx<T> &in, idx<T> &out) {
// allocate padded buffer
idxdim d = in;
d.setdim(1, d.dim(1) + nrow + nrow2);
d.setdim(2, d.dim(2) + ncol + ncol2);
if (out.get_idxdim() != d)
out.resize(d);
idx_clear(out);
// copy in to padded buffer
idx<T> tmp = out.narrow(1, in.dim(1), nrow);
tmp = tmp.narrow(2, in.dim(2), ncol);
idx_copy(in, tmp);
if (bmirror)
mirror(in, out);
}
bool detection_thread<T>::get_data(bboxes &bboxes2, idx<ubyte> &frame2,
uint &total_saved_, std::string &frame_name_,
uint *id, svector<midx<T> > *samples,
bboxes *bbsamples, bool *skipped) {
// lock data
mutex_out.lock();
// only read data if it has been updated
if (!out_updated) {
// unlock data
mutex_out.unlock();
return false;
}
// data is updated, but just to tell we skipped the frame
if (frame_skipped) {
if (skipped) *skipped = true;
frame_skipped = false;
// reset updated flag
out_updated = false;
// declare thread as available
bavailable = true;
// unlock data
mutex_out.unlock();
return false;
}
if (skipped) *skipped = false;
// clear bboxes
bboxes2.clear();
bboxes2.push_back_new(bbs);
bbs.clear(); // no use for local bounding boxes anymore, clear them
// check frame is correctly allocated, if not, allocate.
if (frame2.order() != uframe.order())
frame2 = idx<ubyte>(uframe.get_idxdim());
else if (frame2.get_idxdim() != uframe.get_idxdim())
frame2.resize(uframe.get_idxdim());
// copy frame
idx_copy(uframe, frame2);
// set total of boxes saved
total_saved_ = total_saved;
// set frame name
frame_name_ = frame_name;
// set frame id
if (id) *id = frame_id;
// overwrite samples
if (samples) {
samples->clear();
samples->push_back_new(returned_samples);
returned_samples.clear();
}
if (bbsamples) {
bbsamples->clear();
bbsamples->push_back_new(returned_samples_bboxes);
returned_samples_bboxes.clear();
}
// reset updated flag
out_updated = false;
// declare thread as available
bavailable = true;
// unlock data
mutex_out.unlock();
// confirm that we copied data.
return true;
}
void class_answer<T, Tds1, Tds2>::fprop1(idx<T> &in, idx<T> &out)
{
// resize out if necessary
idxdim d(in);
d.setdim(0, 2); // 2 outputs per pixel: class,confidence
idx<T> outx = out;
idx<T> inx = in;
if (resize_output)
{
if (d != out.get_idxdim())
{
out.resize(d);
outx = out;
}
}
else // if not resizing, narrow to the number of targets
{
if (outx.dim(0) != targets.dim(0))
outx = outx.narrow(0, targets.dim(0), 0);
}
// apply tanh if required
if (apply_tanh)
{
mtanh.fprop1(in, tmp);
inx = tmp;
}
// loop on features (dimension 0) to set class and confidence
int classid;
T conf, max2 = 0;
idx_1loop2(ii, inx, T, oo, outx, T, {
if (binary_target)
{
T t0 = targets.gget(0);
T t1 = targets.gget(1);
T a = ii.gget();
if (std::fabs((double)a - t0) < std::fabs((double)a - t1))
{
oo.set((T)0, 0); // class 0
oo.set((T)(2 - std::fabs((double)a - t0)) / 2, 1); // conf
}
else
{
oo.set((T)1, 0); // class 1
oo.set((T)(2 - std::fabs((double)a - t1)) / 2, 1); // conf
}
}
else if (single_output >= 0)
{
oo.set((T)single_output, 0); // all answers are the same class
oo.set((T)((ii.get(single_output) - target_min) / target_range), 1);
}
else // 1-of-n target
{ // set class answer
if (force_class >= 0) classid = force_class;
else classid = idx_indexmax(ii);
oo.set((T)classid, 0);
// set confidence
intg p;
bool ini = false;
switch (conf_type)
{
case confidence_sqrdist: // squared distance to target
target = targets.select(0, classid);
conf = (T)(1.0 - ((idx_sqrdist(target, ii) - conf_shift)
/ conf_ratio));
oo.set(conf, 1);
break;
case confidence_single: // simply return class' out (normalized)
conf = (T)((ii.get(classid) - conf_shift) / conf_ratio);
oo.set(conf, 1);
break;
case confidence_max: // distance with 2nd max answer
conf = std::max(target_min, std::min(target_max, ii.get(classid)));
for (p = 0; p < ii.dim(0); ++p)
{
if (p != classid)
{
if (!ini)
{
max2 = ii.get(p);
ini = true;
}
else
{
if (ii.get(p) > max2)
max2 = ii.get(p);
}
}
}
max2 = std::max(target_min, std::min(target_max, max2));
oo.set((T)(((conf - max2) - conf_shift) / conf_ratio), 1);
break;
default:
eblerror("confidence type " << conf_type << " undefined");
}
}
});
//! Recursively goes through dir, looking for files matching extension ext.
void process_dir(const char *dir, const char *ext, const char* leftp,
const char *rightp, unsigned int width,
unsigned int fwidth, idx<float> &images,
idx<int> &labels, int label, bool silent, bool display,
bool *binocular,
const int channels_mode, const int channels_size,
idx<ubyte> &classes, unsigned int &counter,
idx<unsigned int> &counters_used, idx<int> &ds_assignment,
unsigned int fkernel_size, int deformations,
idx<int> &deformid, int &ndefid,
idx<ubyte> &ds_names) {
regex eExt(ext);
string el(".*");
idx<float> limg(1, 1, 1);
idx<float> rimg(1, 1, 1);
idx<float> tmp, left_images;
idx<int> current_labels;
idx<int> current_deformid;
idx<float> tmp2;
int current_ds;
unsigned int current_used;
if (leftp) {
el += leftp;
el += ".*";
}
regex eLeft(el);
cmatch what;
path p(dir);
if (!exists(p))
return ;
directory_iterator end_itr; // default construction yields past-the-end
for (directory_iterator itr(p); itr != end_itr; ++itr) {
if (is_directory(itr->status())) {
process_dir(itr->path().string().c_str(), ext, leftp, rightp, width,
fwidth, images, labels, label, silent, display, binocular,
channels_mode, channels_size, classes, counter,
counters_used, ds_assignment, fkernel_size, deformations,
deformid, ndefid, ds_names);
} else if (regex_match(itr->leaf().c_str(), what, eExt)) {
if (regex_match(itr->leaf().c_str(), what, eLeft)) {
current_ds = ds_assignment.get(counter);
if (current_ds != -1) {
current_used = counters_used.get(current_ds);
// found left image
// increase example number
labels.set(label, current_ds, counters_used.get(current_ds));
// check for right image
if (rightp != NULL) {
regex reg(leftp);
string rp(rightp);
string s = regex_replace(itr->leaf(), reg, rp);
string sfull = itr->path().branch_path().string();
sfull += "/";
sfull += s;
path r(sfull);
if (exists(r)) {
// found right image
*binocular = true;
if (image_read_rgbx(r.string().c_str(), rimg)) {
// resize stereo dimension to twice the channels_size
if (images.dim(3) == channels_size)
images.resize(images.dim(0), images.dim(1),
images.dim(2), images.dim(3) * 2);
// take the right most square of the image
if (rimg.dim(0) <= rimg.dim(1))
rimg = rimg.narrow(1, rimg.dim(0),
rimg.dim(1) - rimg.dim(0));
else
rimg = rimg.narrow(0, rimg.dim(1),
rimg.dim(0) - rimg.dim(1));
// resize image to target width
rimg = image_resize(rimg, width, width, 1);
tmp = images[current_ds];
tmp = tmp[counters_used.get(current_ds)];
tmp = tmp.narrow(2, channels_size, channels_size);
// finally copy right images to idx
convert_image(rimg, tmp, channels_mode, fkernel_size);
}
if (!silent)
cout << "Processing (right): " << sfull << endl;
}
}
if (!silent) {
cout << counter << "/" << ds_assignment.dim(0) << ": ";
cout << itr->path().string().c_str() << endl;
}
// process left image
if (image_read_rgbx(itr->path().string().c_str(), limg)) {
// take the left most square of the image
int h = limg.dim(0), w = limg.dim(1), newh, neww;
if (h > w) {
newh = width;
neww = (newh / (float) h) * w;
} else {
neww = width;
newh = (neww / (float) w) * h;
}
// resize image to target width
limg = image_resize(limg, neww, newh, 1);
left_images = images[current_ds];
left_images = left_images.narrow(3, channels_size, 0);
left_images = left_images.narrow(0, 1 + (deformations<=0?0:
deformations),
current_used);
current_labels = labels[current_ds];
current_labels = current_labels.narrow(0, current_labels.dim(0) -
current_used,current_used);
current_deformid = deformid[current_ds];
current_deformid = current_deformid.
narrow(0, current_deformid.dim(0) - current_used, current_used);
tmp = left_images[0];
// convert and copy image into images buffer
convert_image(limg, tmp, channels_mode, fkernel_size);
// if adding to dataset 0, add deformations if deformations > 0
if ((current_ds == 0) && (deformations > 0)) {
add_deformations(left_images, current_labels, label,
deformations, current_deformid, ndefid);
counters_used.set(counters_used.get(current_ds) + deformations,
current_ds);
} else // no deformations
left_images = left_images.narrow(0, 1, 0);
// display
if (display) display_images(classes, label, limg, left_images,
channels_mode, ds_names, current_ds);
// increment counter for dataset current_ds
counters_used.set(counters_used.get(current_ds) + 1, current_ds);
}
}
counter++;
}
}
}
}