autoFeatureWeights FeatureWeights_computeRELIEF
(
///////////////////////////////
// Parameters //
///////////////////////////////
PatternList pp, // source pattern
//
Categories c, // source categories
//
long k // k(!)
//
)
{
autoPatternList p = Data_copy (pp);
autoFeatureWeights me = FeatureWeights_create (p -> nx);
/////////////////////////////////
// Initial weights <- 0 //
/////////////////////////////////
for (long i = 1; i <= p->nx; i++) {
my fweights -> data [1] [i] = 0.0;
}
/////////////////////////////////
// Normalization //
/////////////////////////////////
autoNUMvector <double> min (0L, p->nx - 1);
autoNUMvector <double> max (0L, p->nx - 1);
for (long x = 1; x <= p -> nx; x ++) {
max [x] = p -> z [1] [x]; // BUG: this will just crash because of array index out of bounds
min [x] = max [x];
}
for (long y = 1; y <= p -> ny; y ++) {
for (long x = 1; x <= p->nx; x++) {
if (p->z[y][x] > max[x]) max[x] = p->z[y][x];
if (p->z[y][x] < min[x]) min[x] = p->z[y][x];
}
}
autoNUMvector <double> alfa (0L, p -> nx - 1);
for (long x = 1; x <= p -> nx; x ++) {
alfa [x] = max [x] - min [x]; // BUG: this will just crash because of array index out of bounds
}
for (long y = 1; y <= p->ny; y++) {
for (long x = 1; x <= p->nx; x++) {
if (alfa [x] != 0.0) {
p->z[y][x] = (p->z[y][x] - min[x]) / alfa[x];
} else {
p->z[y][x] = 0.0;
}
}
}
/////////////////////////////////
// Computing prior class probs //
/////////////////////////////////
autoNUMvector <double> priors (0L, c->size - 1); // worst-case allocations
autoNUMvector <long> classes (0L, c->size - 1);//
autoNUMvector <long> enemies (0L, c->size - 1);//
autoNUMvector <long> friends (0L, c->size - 1);//
long nclasses = FeatureWeights_computePriors (c, classes.peek(), priors.peek());
Melder_assert (nclasses >= 2);
/////////////////////////////////
// Updating the w.vector //
/////////////////////////////////
for (long y = 1; y <= p -> ny; y ++) {
long nfriends = KNN_kFriends (p.get(), p.get(), c, y, k, friends.peek());
long nenemies = KNN_kUniqueEnemies (p.get(), p.get(), c, y, nclasses - 1, enemies.peek());
if (nfriends && nenemies) {
autoNUMvector <double> classps (0L, nenemies - 1);
for (long eq = 0; eq < nenemies; eq ++) {
for (long iq = 0; iq < nclasses; iq ++) {
if (FeatureWeights_areFriends (c->at [enemies [eq]], c->at [classes [iq]])) {
classps [eq] = priors [iq];
break;
}
}
}
for (long x = 1; x <= p -> nx; x ++) {
double p1 = 0.0;
double p2 = 0.0;
for (long ec = 0; ec < nfriends; ec ++) {
p1 += fabs (p -> z [y] [x] - p -> z [friends [ec]] [x]) / (p -> ny * nfriends);
}
for (long ec = 0; ec < nenemies; ec++) {
p2 += (fabs (p->z[y][x] - p->z[enemies[ec]][x]) * classps[ec]) / p->ny;
}
my fweights -> data [1] [x] = my fweights -> data [1] [x] - p1 + p2;
}
}
}
return me;
}
bool VLFeat::CalculateCommon(int f, bool all, int l) {
string msg = "VLFeat::CalculateCommon("+ToStr(f)+","+ToStr(all)+","+
ToStr(l)+") : ";
// if (!do_fisher && !do_vlad) {
// cerr << msg
// << "either encoding=fisher or encoding=vlad should be specified"
// << endl;
// return false;
// }
if (!gmm && !kmeans) {
cerr << msg << "either gmm=xxx or kmeans=xxx option should be given"
<< endl;
return false;
}
cox::tictac::func tt(tics, "VLFeat::CalculateCommon");
// obs! only some parameters here, should be in ProcessOptionsAndRemove()
// too, also scales and geometry should be made specifiable...
bool normalizeSift = false, renormalize = true, flat_window = true;
size_t step = 3, binsize = 8;
EnsureImage();
int width = Width(true), height = Height(true);
if (FrameVerbose())
cout << msg+"wxh="
<< width << "x" << height << "=" << width*height << endl;
vector<float> rgbcoeff { 0.2989, 0.5870, 0.1140 };
imagedata idata = CurrentFrame();
idata.convert(imagedata::pixeldata_float);
idata.force_one_channel(rgbcoeff);
vector<float> dsift;
size_t descr_size_orig = 0, descr_size_final = 0;
vector<float> scales { 1.0000, 0.7071, 0.5000, 0.3536, 0.2500 };
// vector<float> scales { 1.0000 };
for (size_t i=0; i<scales.size(); i++) {
if (KeyPointVerbose())
cout << "Starting vl_dsift_process() in scale " << scales[i] << endl;
imagedata simg = idata;
if (scales[i]!=1) {
scalinginfo si(simg.width(), simg.height(),
(int)floor(scales[i]*simg.width()+0.5),
(int)floor(scales[i]*simg.height()+0.5));
simg.rescale(si, 1);
}
// VlDsiftFilter *sf = vl_dsift_new(simg.width(), simg.height());
VlDsiftFilter *sf = vl_dsift_new_basic(simg.width(), simg.height(),
step, binsize);
// opts.scales = logspace(log10(1), log10(.25), 5) ;
// void vl_dsift_set_bounds ( VlDsiftFilter * self,
// int minX,
// int minY,
// int maxX,
// int maxY
// );
// VlDsiftDescriptorGeometry geom = { 8, 4, 4, 0, 0 };
// vl_dsift_set_geometry(sf, &geom);
//vl_dsift_set_steps(sf, 3, 3);
//vl_dsift_set_window_size(sf, 8);
vl_dsift_set_flat_window(sf, flat_window); // aka fast in matlab
vector<float> imgvec = simg.get_float();
const float *img_fp = &imgvec[0];
// cout << "IMAGE = " << img_fp[0] << " " << img_fp[1] << " "
// << img_fp[2] << " ... " << img_fp[41] << endl;
vl_dsift_process(sf, img_fp);
// if opts.rootSift // false
// descrs{si} = sqrt(descrs{si}) ;
// end
// if opts.normalizeSift //true
// descrs{si} = snorm(descrs{si}) ;
// end
descr_size_orig = sf->descrSize;
size_t nf = sf->numFrames;
const VlDsiftKeypoint *k = sf->frames;
float *d = sf->descrs;
if (KeyPointVerbose())
cout << " found " << sf->numFrames << " 'frames' in "
<< simg.info() << endl
<< " descriptor dim " << descr_size_orig << endl;
if (PixelVerbose())
for (size_t i=0; i<nf; i++) {
cout << " i=" << i << " x=" << k[i].x << " y=" << k[i].y
<< " s=" << k[i].s << " norm=" << k[i].norm;
if (FullVerbose()) {
cout << " RAW";
for (size_t j=0; j<descr_size_orig; j++)
cout << " " << d[i*descr_size_orig+j];
}
cout << endl;
}
if (normalizeSift) {
for (size_t i=0; i<nf; i++) {
if (PixelVerbose())
cout << " i=" << i << " x=" << k[i].x << " y=" << k[i].y
<< " s=" << k[i].s << " norm=" << k[i].norm;
double mul = 0.0;
for (size_t j=0; j<descr_size_orig; j++)
mul += d[i*descr_size_orig+j]*d[i*descr_size_orig+j];
if (mul)
mul = 1.0/sqrt(mul);
if (FullVerbose())
cout << " NORM";
for (size_t j=0; j<descr_size_orig; j++) {
d[i*descr_size_orig+j] *= mul;
if (FullVerbose())
cout << " " << d[i*descr_size_orig+j];
}
if (PixelVerbose())
cout << endl;
}
}
if (!pca.vector_length()) {
dsift.insert(dsift.end(), d, d+nf*descr_size_orig);
descr_size_final = descr_size_orig;
} else {
for (size_t i=0; i<nf; i++) {
vector<float> vin(d+i*descr_size_orig, d+(i+1)*descr_size_orig);
vector<float> vout = pca.projection_coeff(vin);
dsift.insert(dsift.end(), vout.begin(), vout.end());
}
descr_size_final = pca.base_size();
}
vl_dsift_delete(sf);
}
size_t numdata = dsift.size()/descr_size_final;
const float *datain = &dsift[0];
vector<float> enc((do_fisher?2:1)*descriptor_dim()*nclusters());
float *dataout = &enc[0];
if (do_fisher) {
if (FrameVerbose())
cout << msg << "fisher encoding " << numdata
<< " descriptors of size " << descr_size_orig
<< " => " << descr_size_final
<< " with gmm dimensionality " << descriptor_dim()
<< endl;
if (descr_size_final!=descriptor_dim()) {
cerr << msg << "dimensionality mismatch descr_size_final="
<< descr_size_final << " descriptor_dim()=" << descriptor_dim()
<< endl;
return false;
}
vl_fisher_encode(dataout, VL_TYPE_FLOAT, means(), descriptor_dim(),
nclusters(), covariances(), priors(),
datain, numdata, VL_FISHER_FLAG_IMPROVED) ;
}
if (do_vlad) {
//obs! correct use of pca?
if (FrameVerbose())
cout << msg << "vlad encoding " << numdata
<< " descriptors of size " << descr_size_final << endl;
vector<vl_uint32> indexes(numdata);
vector<float> distances(numdata);
if (kdtree)
vl_kdforest_query_with_array(kdtree, &indexes[0], 1, numdata, &distances[0], datain);
else
vl_kmeans_quantize(kmeans, &indexes[0], &distances[0], datain, numdata);
vector<float> assignments(numdata*nclusters());
for (size_t i=0; i<numdata; i++)
assignments[i * nclusters() + indexes[i]] = 1;
int vlad_flags = VL_VLAD_FLAG_SQUARE_ROOT|VL_VLAD_FLAG_NORMALIZE_COMPONENTS;
vl_vlad_encode(dataout, VL_TYPE_FLOAT, means(), descriptor_dim(),
nclusters(), datain, numdata, &assignments[0],
vlad_flags);
}
if (renormalize) {
if (PixelVerbose())
cout << " RENORM:";
double mul = 0.0;
for (size_t j=0; j<enc.size(); j++)
mul += enc[j]*enc[j];
if (mul)
mul = 1.0/sqrt(mul);
for (size_t j=0; j<enc.size(); j++) {
if (PixelVerbose())
cout << " " << enc[j];
enc[j] *= mul;
if (PixelVerbose())
cout << "->" << enc[j];
}
if (PixelVerbose())
cout << endl;
}
((VectorData*)GetData(0))->setVector(enc);
return true;
}
