void SelfSimDescriptor::computeLogPolarMapping(Mat& mappingMask) const
{
mappingMask.create(largeSize, largeSize, CV_8S);
// What we want is
// log_m (radius) = numberOfDistanceBuckets
// <==> log_10 (radius) / log_10 (m) = numberOfDistanceBuckets
// <==> log_10 (radius) / numberOfDistanceBuckets = log_10 (m)
// <==> m = 10 ^ log_10(m) = 10 ^ [log_10 (radius) / numberOfDistanceBuckets]
//
int radius = largeSize/2, angleBucketSize = 360 / numberOfAngles;
int fsize = (int)getDescriptorSize();
double inv_log10m = (double)numberOfDistanceBuckets/log10((double)radius);
for (int y=-radius ; y<=radius ; y++)
{
schar* mrow = mappingMask.ptr<schar>(y+radius);
for (int x=-radius ; x<=radius ; x++)
{
int index = fsize;
float dist = (float)std::sqrt((float)x*x + (float)y*y);
int distNo = dist > 0 ? cvRound(log10(dist)*inv_log10m) : 0;
if( startDistanceBucket <= distNo && distNo < numberOfDistanceBuckets )
{
float angle = std::atan2( (float)y, (float)x ) / (float)CV_PI * 180.0f;
if (angle < 0) angle += 360.0f;
int angleInt = (cvRound(angle) + angleBucketSize/2) % 360;
int angleIndex = angleInt / angleBucketSize;
index = (distNo-startDistanceBucket)*numberOfAngles + angleIndex;
}
mrow[x + radius] = saturate_cast<schar>(index);
}
}
}
void BlockCellRemapper::map(std::vector< float >& block_feats)
{
assert(block_feats.size() == computer.getDescriptorSize());
vector<float> cell_feats(getDescriptorSize(),0);
// getIndex(int blockX, int blockY, int cellN, int bin);
assert(computer.cellsPerBlock() == 4);
for(int blockX = 0; blockX < computer.blocks_x(); blockX++)
for(int blockY = 0; blockY < computer.blocks_y(); blockY++)
for(int cellN = 0; cellN < 4; cellN++)
for(int bin = 0; bin < computer.getNBins(); bin++)
{
// get from the celly form (cells)
// notes
// dst block depends on cell # and src block
int cell_cell_X, cell_cell_Y;
map_blockToCell4(blockX, blockY, cellN, cell_cell_X, cell_cell_Y);
// figure out where to put it in the dest
int icell = getIndex(cell_cell_X,cell_cell_Y,0,bin+cellN*computer.getNBins());
assert(icell >= 0 && icell < cell_feats.size());
float&cell_val = cell_feats[icell];
// get each piece of data from in the blocky
int iblk = computer.getIndex(blockX,blockY,cellN,bin);
assert(iblk >= 0 && iblk < block_feats.size());
float&blk_val = block_feats[iblk];
// do the copy
cell_val = blk_val;
}
block_feats = cell_feats;
}
// Cell Format to block format
void BlockCellRemapper::unmap(std::vector< double >& cell_feats)
{
if(cell_feats.size() != getDescriptorSize())
{
cout << printfpp("%d != %d",cell_feats.size(),getDescriptorSize()) << endl;
throw std::logic_error("f**k!!!");
}
vector<double> block_feats(computer.getDescriptorSize(),0);
// getIndex(int blockX, int blockY, int cellN, int bin);
assert(computer.cellsPerBlock() == 4);
for(int blockX = 0; blockX < computer.blocks_x(); blockX++)
for(int blockY = 0; blockY < computer.blocks_y(); blockY++)
for(int cellN = 0; cellN < 4; cellN++)
for(int bin = 0; bin < computer.getNBins(); bin++)
{
// get from the celly form (cells)
// notes
// dst block depends on cell # and src block
int cell_cell_X, cell_cell_Y;
map_blockToCell4(blockX, blockY, cellN, cell_cell_X, cell_cell_Y);
// figure out where to put it in the dest
int icell = getIndex(cell_cell_X,cell_cell_Y,0,bin+cellN*computer.getNBins());
if(!(icell >= 0 && icell < cell_feats.size()))
{
printf("cellX = %d cellY = %d\n",cell_cell_X,cell_cell_Y);
printf("icell = %d of %d\n",icell,(int)cell_feats.size());
}
assert(icell >= 0 && icell < cell_feats.size());
double&cell_val = cell_feats[icell];
// get each piece of data from in the blocky
int iblk = computer.getIndex(blockX,blockY,cellN,bin);
assert(iblk >= 0 && iblk < block_feats.size());
double&blk_val = block_feats[iblk];
// do the copy
blk_val = cell_val;
}
cell_feats = block_feats;
}
cl_int IntelAccelerator::getInfo(cl_accelerator_info_intel paramName,
size_t paramValueSize,
void *paramValue,
size_t *paramValueSizeRet) const {
cl_int result = CL_SUCCESS;
size_t ret = 0;
switch (paramName) {
case CL_ACCELERATOR_DESCRIPTOR_INTEL: {
ret = getDescriptorSize();
result = ::getInfo(paramValue, paramValueSize, getDescriptor(), ret);
}
break;
case CL_ACCELERATOR_REFERENCE_COUNT_INTEL: {
auto v = getReference();
ret = sizeof(cl_uint);
result = ::getInfo(paramValue, paramValueSize, &v, ret);
}
break;
case CL_ACCELERATOR_CONTEXT_INTEL: {
ret = sizeof(cl_context);
cl_context ctx = static_cast<cl_context>(pContext);
result = ::getInfo(paramValue, paramValueSize, &ctx, ret);
}
break;
case CL_ACCELERATOR_TYPE_INTEL: {
auto v = getTypeId();
ret = sizeof(cl_accelerator_type_intel);
result = ::getInfo(paramValue, paramValueSize, &v, ret);
}
break;
default:
result = CL_INVALID_VALUE;
break;
}
if (paramValueSizeRet) {
*paramValueSizeRet = ret;
}
return result;
}
void FeatureBinCombiner::compute(const ImRGBZ& im, std::vector< float >& feats)
{
// check the depth image for NaNs
//assert(!any<float>(im.Z,[](float v){return !goodNumber(v);}));
bool rows_match = im.rows() == blocks_y()*getCellSize().height;
bool cols_match = im.cols() == blocks_x()*getCellSize().width;
if(!(rows_match && cols_match))
{
cout << printfpp("imSize = %d %d",im.cols(),im.rows()) << endl;
cout << printfpp("blocks_x[%d]*cell_size.width[%d] = %d",
blocks_x(),getCellSize().width,blocks_x()*getCellSize().width) << endl;
cout << printfpp("blocks_y[%d]*cell_size.height[%d] = %d",
blocks_y(),getCellSize().height,blocks_y()*getCellSize().height) << endl;
assert(rows_match && cols_match);
}
// precompute the features with each computer.
vector<vector<float> > in_feats;
int iter = 0;
for(shared_ptr<DepthFeatComputer>&computer : computers)
{
in_feats.push_back(vector<float>());
computer->compute(im,in_feats.back());
// check for errors
for(float&value : in_feats.back())
{
bool ok = goodNumber(value);
if(!ok)
cout << "problem @ computer = " << iter << endl;
assert(ok);
}
iter++;
}
// combine into output
feats = vector<float>(getDescriptorSize(),0);
for(int blockX = 0; blockX < blocks_x(); blockX++)
for(int blockY = 0; blockY < blocks_y(); blockY++)
{
int outbin = 0;
for(int compIter = 0; compIter < computers.size(); compIter++)
{
shared_ptr<DepthFeatComputer> computer = computers[compIter];
assert(computer->cellsPerBlock() == 1);
for(int inbin = 0; inbin < computer->getNBins(); inbin++,outbin++)
feats[getIndex(blockX,blockY,0,outbin)] =
in_feats[compIter][computer->getIndex(blockX,blockY,0,inbin)]*weights[compIter];
}
}
}
/// pack into a form [ori1 ori2 ori .... oriN norm1 ... norm4]
void HOGComputer18p4_General::compute(const ImRGBZ& im, std::vector< float >& feats)
{
vector<float> unreduced; hog18x4mapped.compute(im,unreduced);
int raw_bins = hog18x4mapped.getNBins();
int nbins = params::ORI_BINS;
assert(raw_bins == nbins*4);
/// now, iterate over all the cells in the remapped
// 18x4 feature, reduce the block, and write the reduced
// form to the 18p4 feature.
feats = vector<float>(getDescriptorSize(),0);
for(int yIter = 0; yIter < blocks_y(); yIter++)
for(int xIter = 0; xIter < blocks_x(); xIter++)
{
// (1) compute the orientation sum
for(int out_binIter = 0; out_binIter < nbins; out_binIter++)
{
float ori_sum = 0;
for(int in_block = 0; in_block < 4; in_block++)
{
int bin_id = out_binIter + nbins*in_block;
ori_sum += unreduced[hog18x4mapped.getIndex(xIter,yIter,0,bin_id)];
}
feats[getIndex(xIter,yIter,0,out_binIter)] = .5*ori_sum;
}
// (2) compute the normalization sum
for(int out_blockIter = 0; out_blockIter < 4; out_blockIter++)
{
float block_sum = 0;
for(int in_oriBin = 0; in_oriBin < nbins; in_oriBin++)
{
int bin_id = in_oriBin + out_blockIter*nbins;
block_sum += unreduced[hog18x4mapped.getIndex(xIter,yIter,0,bin_id)];
}
feats[getIndex(xIter,yIter,0,nbins + out_blockIter)] = .2357*block_sum;
}
}
}
void PCAFeature::compute(const ImRGBZ& im, std::vector< float >& reduced_feats)
{
// allocate space
vector<float> unreduced_feats; unreduced_computer.compute(im,unreduced_feats);
// wrap with matrices
int ncells = blocks_x()*blocks_y();
Mat mat_unreduced_feats(
ncells,unreduced_computer.getNBins(),DataType<float>::type,&unreduced_feats[0]);
// compute the PCA
Mat mat_reduced_feats = g_pca_reducer->project(mat_unreduced_feats);
// must I now divide by the eigenvalues?
assert(g_pca_reducer->eigenvalues.type() == DataType<float>::type);
for(int rIter = 0; rIter < mat_reduced_feats.rows; rIter++)
for(int cIter = 0; cIter < mat_reduced_feats.cols; cIter++)
mat_reduced_feats.at<float>(rIter,cIter) /= g_pca_reducer->eigenvalues.at<float>(cIter);
// convert to a flat feature vector
reduced_feats = mat_reduced_feats.reshape(0,mat_reduced_feats.size().area());
assert(reduced_feats.size() == getDescriptorSize());
}
void SelfSimDescriptor::compute(const Mat& img, vector<float>& descriptors, Size winStride,
const vector<Point>& locations) const
{
CV_Assert( img.depth() == CV_8U );
winStride.width = std::max(winStride.width, 1);
winStride.height = std::max(winStride.height, 1);
Size gridSize = getGridSize(img.size(), winStride);
int i, nwindows = locations.empty() ? gridSize.width*gridSize.height : (int)locations.size();
int border = largeSize/2 + smallSize/2;
int fsize = (int)getDescriptorSize();
vector<float> tempFeature(fsize+1);
descriptors.resize(fsize*nwindows + 1);
Mat ssd(largeSize, largeSize, CV_32F), mappingMask;
computeLogPolarMapping(mappingMask);
#if 0 //def _OPENMP
int nthreads = cvGetNumThreads();
#pragma omp parallel for num_threads(nthreads)
#endif
for( i = 0; i < nwindows; i++ )
{
Point pt;
float* feature0 = &descriptors[fsize*i];
float* feature = &tempFeature[0];
int x, y, j;
if( !locations.empty() )
{
pt = locations[i];
if( pt.x < border || pt.x >= img.cols - border ||
pt.y < border || pt.y >= img.rows - border )
{
for( j = 0; j < fsize; j++ )
feature0[j] = 0.f;
continue;
}
}
else
pt = Point((i % gridSize.width)*winStride.width + border,
(i / gridSize.width)*winStride.height + border);
SSD(img, pt, ssd);
// Determine in the local neighborhood the largest difference and use for normalization
float var_noise = 1000.f;
for( y = -1; y <= 1 ; y++ )
for( x = -1 ; x <= 1 ; x++ )
var_noise = std::max(var_noise, ssd.at<float>(largeSize/2+y, largeSize/2+x));
for( j = 0; j <= fsize; j++ )
feature[j] = FLT_MAX;
// Derive feature vector before exp(-x) computation
// Idea: for all x,a >= 0, a=const. we have:
// max [ exp( -x / a) ] = exp ( -min(x) / a )
// Thus, determine min(ssd) and store in feature[...]
for( y = 0; y < ssd.rows; y++ )
{
const schar *mappingMaskPtr = mappingMask.ptr<schar>(y);
const float *ssdPtr = ssd.ptr<float>(y);
for( x = 0 ; x < ssd.cols; x++ )
{
int index = mappingMaskPtr[x];
feature[index] = std::min(feature[index], ssdPtr[x]);
}
}
var_noise = -1.f/var_noise;
for( j = 0; j < fsize; j++ )
feature0[j] = feature[j]*var_noise;
Mat _f(1, fsize, CV_32F, feature0);
cv::exp(_f, _f);
}
}
