//! Find the nearest neighbors in the descriptor space using FLANN.
void append_nearest_neighbors(
size_t i1,
const Set<OERegion, RealDescriptor>& keys1,
const Set<OERegion, RealDescriptor>& keys2,
vector<Match>& matches,
const flann::Matrix<float>& /*data2*/,
flann::Index<flann::L2<float>>& tree2,
float squared_ratio_thres,
Match::Direction dir,
bool self_matching, const KeyProximity& is_redundant, // self-matching
vector<int>& vec_indices,
vector<float>& vec_dists,
size_t num_max_neighbors) // internal storage parameters
{
// Prepare the query matrix
flann::Matrix<float> query{
const_cast<float *>(&(keys1.descriptors.matrix()(0, i1))),
1, keys1.descriptors.dimension()
};
// Prepare the indices and distances.
flann::Matrix<int> indices{ &vec_indices[0], 1, num_max_neighbors };
flann::Matrix<float> dists{ &vec_dists[0], 1, num_max_neighbors };
// Create search parameters.
flann::SearchParams search_params;
// N.B.: We should not be in the boundary case in practice, in which case the
// ambiguity score does not really make sense.
//
// Boundary case 1.
if (keys2.size() == 0)
return;
// Boundary case 2.
if (keys2.size() == 1 && !self_matching)
{
auto m = Match{
&keys1.features[i1], &keys2.features[0], 1.f, dir, int(i1), 0
};
m.rank() = 1;
if(dir == Match::Direction::TargetToSource)
{
swap(m.x_pointer(), m.y_pointer());
swap(m.x_index(), m.y_index());
}
if (m.score() < squared_ratio_thres)
matches.push_back(m);
return;
}
// Boundary case 3.
if (keys2.size() == 2 && self_matching)
{
tree2.knnSearch(query, indices, dists, 2, search_params);
const auto i2 = indices[0][1]; // The first index can't be indices[0][0], which is i1.
auto m = Match{
&keys1.features[i1], &keys2.features[i2], 1.f, dir,
int(i1), i2
};
m.rank() = 1;
if(dir == Match::Direction::TargetToSource)
{
swap(m.x_pointer(), m.y_pointer());
swap(m.x_index(), m.y_index());
}
if (m.score() < squared_ratio_thres)
matches.push_back(m);
return;
}
// Now treat the generic case.
//
// Search the nearest neighbor.
tree2.knnSearch(query, indices, dists, 3, search_params);
// This is to avoid the source key matches with himself in case of intra
// image matching.
const auto top1_index = self_matching ? 1 : 0;
auto top1_score = dists[0][top1_index + 1] > 0.f ?
dists[0][top1_index] / dists[0][top1_index + 1] : 0.f;
auto K = 1;
// Determine the number of nearest neighbors.
if (squared_ratio_thres > 1.f)
{
// Performs an adaptive radius search.
const auto radius = dists[0][top1_index] * squared_ratio_thres;
K = tree2.radiusSearch(query, indices, dists, radius, search_params);
}
// Retrieve the right key points.
for (int rank = top1_index; rank < K; ++rank)
{
auto score = 0.f;
if (rank == top1_index)
score = top1_score;
else if (dists[0][top1_index])
score = dists[0][rank] / dists[0][top1_index];
// We still need this check as FLANN can still return wrong neighbors.
if (score > squared_ratio_thres)
break;
auto i2 = indices[0][rank];
// Ignore the match if keys1 == keys2.
if (self_matching && is_redundant(keys1.features[i1], keys2.features[i2]))
continue;
Match m(&keys1.features[i1], &keys2.features[i2], score, dir, i1, i2);
m.rank() = top1_index == 0 ? rank+1 : rank;
if(dir == Match::Direction::TargetToSource)
{
swap(m.x_pointer(), m.y_pointer());
swap(m.x_index(), m.y_index());
}
matches.push_back(m);
}
}
// x座標列を返す
const std::vector<double>& FieldBase::x_list() const{
return *x_pointer();
}