我试图在项目中使用nanoflann并正在查看矢量矢量和半径搜索示例。
我找不到使用与坐标类型不同的数据类型执行半径搜索的方法。例如，我的坐标是vector的{{1}};我试图输入uint8_t类型的半径，但收效甚微。
我在源代码中看到uint32_t（我用于距离）使用带有两个模板参数的metric_L2 struct。 L2_Adaptor本身有三个参数，第三个默认为第一个，如果我正确理解代码，这似乎是个问题。但是，尝试强制使用第三个总是会在半径搜索中产生0个匹配。
有办法做到这一点吗？
编辑：在下面的相同代码中，一切正常。但是，如果我将search_radius（和ret_matches）更改为uint32_t，则radiusSearch方法不起作用。
L2_Adaptor更多信息：所以似乎距离类型（#include <iostream>
#include <Eigen/Dense>
#include <nanoflann.hpp>
typedef Eigen::Matrix<uint8_t, Eigen::Dynamic, 1> coord_t;
using namespace nanoflann;
struct Point
{
coord_t address;
Point() {}
Point(uint8_t coordinates) : address(coord_t::Random(coordinates)) {}
};
struct Container
{
std::vector<Point> points;
Container(uint8_t coordinates, uint32_t l)
: points(l)
{
for(auto& each_location: points)
{
each_location = Point(coordinates);
}
}
};
struct ContainerAdaptor
{
typedef ContainerAdaptor self_t;
typedef nanoflann::metric_L2::traits<uint8_t, self_t>::distance_t metric_t;
typedef KDTreeSingleIndexAdaptor<metric_t, self_t, -1, size_t> index_t;
index_t *index;
const Container &container;
ContainerAdaptor(const int dimensions, const Container &container, const int leaf_max_size = 10)
: container(container)
{
assert(container.points.size() != 0 && container.points[0].address.rows() != 0);
const size_t dims = container.points[0].address.rows();
index = new index_t(dims, *this, nanoflann::KDTreeSingleIndexAdaptorParams(leaf_max_size));
index->buildIndex();
}
~ContainerAdaptor()
{
delete index;
}
inline void query(const uint8_t *query_point, const size_t num_closest, size_t *out_indices, uint32_t *out_distances_sq, const int ignoreThis = 10) const
{
nanoflann::KNNResultSet<uint32_t, size_t, size_t> resultSet(num_closest);
resultSet.init(out_indices, out_distances_sq);
index->findNeighbors(resultSet, query_point, nanoflann::SearchParams());
}
const self_t& derived() const
{
return *this;
}
self_t& derived()
{
return *this;
}
inline size_t kdtree_get_point_count() const
{
return container.points.size();
}
inline size_t kdtree_distance(const uint8_t *p1, const size_t idx_p2, size_t size) const
{
size_t s = 0;
for (size_t i = 0; i < size; i++)
{
const uint8_t d = p1[i] - container.points[idx_p2].address[i];
s += d * d;
}
return s;
}
inline coord_t::Scalar kdtree_get_pt(const size_t idx, int dim) const
{
return container.points[idx].address[dim];
}
template <class BBOX>
bool kdtree_get_bbox(BBOX & bb) const
{
for(size_t i = 0; i < bb.size(); i++)
{
bb[i].low = 0;
bb[i].high = UINT8_MAX;
}
return true;
}
};
void container_demo(const size_t points, const size_t coordinates)
{
Container s(coordinates, points);
coord_t query_pt(coord_t::Random(coordinates));
typedef ContainerAdaptor my_kd_tree_t;
my_kd_tree_t mat_index(coordinates, s, 25);
mat_index.index->buildIndex();
const uint8_t search_radius = static_cast<uint8_t>(100);
std::vector<std::pair<size_t, uint8_t>> ret_matches;
nanoflann::SearchParams params;
const size_t nMatches = mat_index.index->radiusSearch(query_pt.data(), search_radius, ret_matches, params);
for (size_t i = 0; i < nMatches; i++)
{
std::cout << "idx[" << i << "]=" << +ret_matches[i].first << " dist[" << i << "]=" << +ret_matches[i].second << std::endl;
}
std::cout << std::endl;
std::cout << "radiusSearch(): radius=" << +search_radius << " -> " << +nMatches << " matches" << std::endl;
}
int main()
{
container_demo(1e6, 32);
return 0;
}
的第三个参数）必须是带符号的类型。如果L2_Adaptor和search_radius也更改为int64_t，则将metric_t typedef更改为以下内容即可解决问题。
ret_matches