Flann Manual 1.8.4

%E8%81%9A%E7%B1%BB%E5%B7%A5%E5%85%B7flann_manual-1.8.4

flann_manual-1.8.4

manual

User Manual:

Open the PDF directly: View PDF .
Page Count: 29

Download
Open PDF In Browser	View PDF

FLANN - Fast Library for Approximate Nearest
Neighbors
User Manual

Marius Muja, mariusm@cs.ubc.ca
David Lowe, lowe@cs.ubc.ca
January 24, 2013

1

Introduction

We can define the nearest neighbor search (NSS) problem in the following way:
given a set of points P = p1 , p2 , . . . , pn in a metric space X, these points must
be preprocessed in such a way that given a new query point q ∈ X, finding the
point in P that is nearest to q can be done quickly.
The problem of nearest neighbor search is one of major importance in a
variety of applications such as image recognition, data compression, pattern
recognition and classification, machine learning, document retrieval systems,
statistics and data analysis. However, solving this problem in high dimensional
spaces seems to be a very difficult task and there is no algorithm that performs
significantly better than the standard brute-force search. This has lead to an
increasing interest in a class of algorithms that perform approximate nearest
neighbor searches, which have proven to be a good-enough approximation in
most practical applications and in most cases, orders of magnitude faster that
the algorithms performing the exact searches.
FLANN (Fast Library for Approximate Nearest Neighbors) is a library for
performing fast approximate nearest neighbor searches. FLANN is written in
the C++ programming language. FLANN can be easily used in many contexts
through the C, MATLAB and Python bindings provided with the library.

1.1

Quick Start

This section contains small examples of how to use the FLANN library from
different programming languages (C++, C, MATLAB and Python).
• C++
// file flann_example.cpp
#include 
#include 
#include 
int main(int argc, char** argv)
{
int nn = 3;
flann::Matrix dataset;
flann::Matrix query;
flann::load_from_file(dataset, "dataset.hdf5","dataset");
flann::load_from_file(query, "dataset.hdf5","query");
flann::Matrix indices(new int[query.rows*nn], query.rows, nn);
flann::Matrix dists(new float[query.rows*nn], query.rows, nn);
// construct an randomized kd-tree index using 4 kd-trees
flann::Index > index(dataset, flann::KDTreeIndexParams(4));
index.buildIndex();
// do a knn search, using 128 checks
index.knnSearch(query, indices, dists, nn, flann::SearchParams(128));
flann::save_to_file(indices,"result.hdf5","result");

1

delete[]
delete[]
delete[]
delete[]

dataset.ptr();
query.ptr();
indices.ptr();
dists.ptr();

return 0;
}

• C
/* file flann_example.c */
#include "flann.h"
#include 
#include 
/* Function that reads a dataset */
float* read_points(char* filename, int *rows, int *cols);
int main(int argc, char** argv)
{
int rows,cols;
int t_rows, t_cols;
float speedup;
/* read dataset points from file dataset.dat */
float* dataset = read_points("dataset.dat", &rows, &cols);
float* testset = read_points("testset.dat", &t_rows, &t_cols);
/* points in dataset and testset should have the same dimensionality */
assert(cols==t_cols);
/* number of nearest neighbors to search */
int nn = 3;
/* allocate memory for the nearest-neighbors indices */
int* result = (int*) malloc(t_rows*nn*sizeof(int));
/* allocate memory for the distances */
float* dists = (float*) malloc(t_rows*nn*sizeof(float));
/* index parameters are stored here */
struct FLANNParameters p = DEFAULT_FLANN_PARAMETERS;
p.algorithm = FLANN_INDEX_AUTOTUNED; /* or FLANN_INDEX_KDTREE, FLANN_INDEX_KMEANS, ... /*
p.target_precision = 0.9; /* want 90% target precision */
/* compute the 3 nearest-neighbors of each point in the testset */
flann_find_nearest_neighbors(dataset, rows, cols, testset, t_rows,
result, dists, nn, &p);
...
free(dataset);
free(testset);
free(result);
free(dists);
return 0;
}

• MATLAB
% create random dataset and test set
dataset = single(rand(128,10000));
testset = single(rand(128,1000));
% define index and search parameters

2

params.algorithm = ’kdtree’;
params.trees = 8;
params.checks = 64;
% perform the nearest-neighbor search
[result, dists] = flann_search(dataset,testset,5,params);

• Python
from pyflann import *
from numpy import *
from numpy.random import *
dataset = rand(10000, 128)
testset = rand(1000, 128)
flann = FLANN()
result,dists = flann.nn(dataset,testset,5,algorithm="kmeans",
branching=32, iterations=7, checks=16);

2

Downloading and compiling FLANN

FLANN can be downloaded from the following address:
http://www.cs.ubc.ca/∼mariusm/flann
After downloading and unpacking, the following files and directories should
be present:
• bin: directory various for scripts and binary files
• doc: directory containg this documentation
• examples: directory containg examples of using FLANN
• src: directory containg the source files
• test: directory containg unit tests for FLANN
To compile the FLANN library the CMake 1 build system is required. Below
is an example of how FLANN can be compiled on Linux (replace x.y.z with the
corresponding version number).
$
$
$
$
$

cd flann-x.y.z-src
mkdir build
cd build
cmake ..
make

On windows the steps are very similar:
1 http://www.cmake.org/

3

>
>
>
>
>
>

"C:\Program Files\Microsoft Visual Studio 9.0\VC\vcvarsall.bat"
cd flann-x.y.z-src
mkdir build
cd build
cmake ..
nmake

There are several compile options that can be configured before FLANN
is compiled, for example the build type (Release, RelWithDebInfo, Debug) or
whether to compile the C, Python or the MATLAB bindings. To change any of
this options use the cmake-gui application after cmake has finished (see figure
1).
> cmake-gui .

Figure 1: Configuring the FLANN compile options

4

2.1

Upgrading from a previous version

This section contains changes that you need to be aware of when upgrading
from a previous version of FLANN.
Version 1.8.0 Due to changes in the library, the on-disk format of the saved
indexes has changed and it is not possible to load indexes previously saved
with an older version of the library.
Version 1.7.0 A small breaking API change in flann::Matrix requires client
code to be updated. In order to release the memory used by a matrix,
use:
delete[] matrix.ptr();

instead of:
delete[] matrix.data;

or:
matrix.free();

The member data of flann::Matrix is not publicly accessible any more,
use the ptr() method to obtain the pointer to the memory buffer.

2.2

Compiling FLANN with multithreading support

For taking advantage of multithreaded search, the project that uses FLANN
needs to be compiled with a compiler that supports the OpenMP standard and
the OpenMP support must be enabled. The number of cores to be used can
be selected with the cores field in the SearchParams structure. By default a
single core will be used. Setting the cores field to zero will automatically use
as many threads as cores available on the machine.

3
3.1

Using FLANN
Using FLANN from C++

The core of the FLANN library is written in C++. To make use of the full
power and flexibility of the templated code one should use the C++ bindings
if possible. To use the C++ bindings you only need to include the the library
header file flann.hpp. An example of the compile command that must be used
will look something like this:
g++ flann_example.cpp -I $FLANN_ROOT/include -o flann_example_cpp

where $FLANN ROOT is the library main directory.
The following sections describe the public C++ API.

5

3.1.1

flann::Index

The FLANN nearest neighbor index class. This class is used to abstract different
types of nearest neighbor search indexes. The class is templated on the distance
functor to be used for computing distances between pairs of features.
namespace flann
{
template
class Index
{
typedef typename Distance::ElementType ElementType;
typedef typename Distance::ResultType DistanceType;
public:
Index(const IndexParams& params, Distance distance = Distance() );
Index(const Matrix& points, const IndexParams& params,
Distance distance = Distance() );
~Index();
void buildIndex();
void buildIndex(const Matrix& points);
void addPoints(const Matrix& points,
float rebuild_threshold = 2);
void removePoint(size_t point_id);
ElementType* getPoint(size_t point_id);
int knnSearch(const Matrix& queries,
Matrix& indices,
Matrix& dists,
size_t knn,
const SearchParams& params);
int knnSearch(const Matrix& queries,
std::vector< std::vector >& indices,
std::vector >& dists,
size_t knn,
const SearchParams& params);
int radiusSearch(const Matrix& queries,
Matrix& indices,
Matrix& dists,
float radius,
const SearchParams& params);
int radiusSearch(const Matrix& queries,
std::vector< std::vector >& indices,
std::vector >& dists,
float radius,
const SearchParams& params);
void save(std::string filename);

6

int veclen() const;
int size() const;
IndexParams getParameters() const;
flann_algorithm_t getType() const;
};
}

The Distance functor
The distance functor is a class whose operator() computes the distance
between two features. If the distance is also a kd-tree compatible distance
it should also provide an accum dist() method that computes the distance
between individual feature dimensions. A typical distance functor looks like
this (see the dist.h file for more examples):
template
struct L2
{
typedef bool is_kdtree_distance;
typedef T ElementType;
typedef typename Accumulator::Type ResultType;
template 
ResultType operator()(Iterator1 a, Iterator2 b, size_t size,
ResultType /*worst_dist*/ = -1) const
{
ResultType result = ResultType();
ResultType diff;
for(size_t i = 0; i < size; ++i ) {
diff = *a++ - *b++;
result += diff*diff;
}
return result;
}
template 
inline ResultType accum_dist(const U& a, const V& b, int) const
{
return (a-b)*(a-b);
}
};

In addition to operator() and accum dist(), a distance functor should
also define the ElementType and the ResultType as the types of the elements
it operates on and the type of the result it computes.
If a distance functor can be used as a kd-tree distance (meaning that the full
distance between a pair of features can be accumulated from the partial distances
between the individual dimensions) a typedef is kdtree distance should be
7

present inside the distance functor. If the distance is not a kd-tree distance, but
it’s a distance in a vector space (the individual dimensions of the elements it operates on can be accessed independently) a typedef is vector space distance
should be defined inside the functor. If neither typedef is defined, the distance
is assumed to be a metric distance and will only be used with indexes operating
on generic metric distances.
flann::Index::Index Constructs a nearest neighbor search index for a given
dataset.
Index(const IndexParams& params, Distance distance = Distance() );
Index(const Matrix& points, const IndexParams& params,
Distance distance = Distance() );

features Matrix containing the features(points) that should be indexed, stored
in a row-major order (one point on each row of the matrix). The size of
the matrix is num f eatures × dimensionality.
params Structure containing the index parameters. The type of index that
will be constructed depends on the type of this parameter. The possible
parameter types are:
LinearIndexParams When passing an object of this type, the index will
perform a linear, brute-force search.
struct LinearIndexParams : public IndexParams
{
};

KDTreeIndexParams When passing an object of this type the index
constructed will consist of a set of randomized kd-trees which will be
searched in parallel.
struct KDTreeIndexParams : public IndexParams
{
KDTreeIndexParams( int trees = 4 );
};

trees The number of parallel kd-trees to use. Good values are in the
range [1..16]
KMeansIndexParams When passing an object of this type the index
constructed will be a hierarchical k-means tree.
struct KMeansIndexParams : public IndexParams
{
KMeansIndexParams( int branching = 32,
int iterations = 11,
flann_centers_init_t centers_init = FLANN_CENTERS_RANDOM,
float cb_index = 0.2 );
};

8

branching The branching factor to use for the hierarchical k-means tree
iterations The maximum number of iterations to use in the k-means
clustering stage when building the k-means tree. If a value of -1 is
used here, it means that the k-means clustering should be iterated
until convergence
centers init The algorithm to use for selecting the initial centers when
performing a k-means clustering step. The possible values are CENTERS RANDOM (picks the initial cluster centers randomly), CENTERS GONZALES (picks the initial centers using Gonzales’ algorithm) and CENTERS KMEANSPP (picks the initial centers using
the algorithm suggested in [AV07])
cb index This parameter (cluster boundary index) influences the way
exploration is performed in the hierarchical kmeans tree. When
cb index is zero the next kmeans domain to be explored is choosen
to be the one with the closest center. A value greater then zero also
takes into account the size of the domain.
CompositeIndexParams When using a parameters object of this type
the index created combines the randomized kd-trees and the hierarchical
k-means tree.
struct CompositeIndexParams : public IndexParams
{
CompositeIndexParams( int trees = 4,
int branching = 32,
int iterations = 11,
flann_centers_init_t centers_init = FLANN_CENTERS_RANDOM,
float cb_index = 0.2 );
};

KDTreeSingleIndexParams When passing an object of this type the
index will contain a single kd-tree optimized for searching lower dimensionality data (for example 3D point clouds)
struct KDTreeSingleIndexParams : public IndexParams
{
KDTreeSingleIndexParams( int leaf_max_size = 10 );
};

max leaf size The maximum number of points to have in a leaf for not
branching the tree any more.
KDTreeCuda3dIndexParams When passing an object of this type the
index will be a single kd-tree that is built and performs searches on a
CUDA compatible GPU. Search performance is best for large numbers of
search and query points. For more information, see section 3.1.10
struct KDTreeCuda3dIndexParams : public IndexParams
{
KDTreeCuda3dIndexParams( int leaf_max_size = 64 );
};

9

max leaf size The maximum number of points to have in a leaf for not
branching the tree any more.
HierarchicalClusteringIndexParams When passing an object of this
type the index constructed will be a hierarchical clustering index. This
type of index works with any metric distance and can be used for matching
binary features using Hamming distances.
struct HierarchicalClusteringIndexParams : public IndexParams
{
HierarchicalClusteringIndexParams(int branching = 32,
flann_centers_init_t centers_init = FLANN_CENTERS_RANDOM,
int trees = 4, int leaf_max_size = 100)
};

branching The branching factor to use for the hierarchical clustering tree
centers init The algorithm to use for selecting the initial centers when
performing a k-means clustering step. The possible values are CENTERS RANDOM (picks the initial cluster centers randomly), CENTERS GONZALES (picks the initial centers using Gonzales’ algorithm) and CENTERS KMEANSPP (picks the initial centers using
the algorithm suggested in [AV07])
trees The number of parallel trees to use. Good values are in the range
[3..8]
leaf size The maximum number of points a leaf node should contain.
LshIndexParams When passing an object of this type the index constructed will be a multi-probe LSH (Locality-Sensitive Hashing) index.
This type of index can only be used for matching binary features using
Hamming distances.
struct LshIndexParams : public IndexParams
{
LshIndexParams(unsigned int table_number = 12,
unsigned int key_size = 20,
unsigned int multi_probe_level = 2);
};

table number The number of hash tables to use
key size The length of the key in the hash tables
multi probe level Number of levels to use in multi-probe (0 for standard
LSH)
AutotunedIndexParams When passing an object of this type the index
created is automatically tuned to offer the best performance, by choosing
the optimal index type (randomized kd-trees, hierarchical kmeans, linear)
and parameters for the dataset provided.

10

struct AutotunedIndexParams : public IndexParams
{
AutotunedIndexParams( float target_precision = 0.9,
float build_weight = 0.01,
float memory_weight = 0,
float sample_fraction = 0.1 );
};

target precision Is a number between 0 and 1 specifying the percentage
of the approximate nearest-neighbor searches that return the exact
nearest-neighbor. Using a higher value for this parameter gives more
accurate results, but the search takes longer. The optimum value
usually depends on the application.
build weight Specifies the importance of the index build time raported
to the nearest-neighbor search time. In some applications it’s acceptable for the index build step to take a long time if the subsequent
searches in the index can be performed very fast. In other applications it’s required that the index be build as fast as possible even if
that leads to slightly longer search times.
memory weight Is used to specify the tradeoff between time (index
build time and search time) and memory used by the index. A value
less than 1 gives more importance to the time spent and a value
greater than 1 gives more importance to the memory usage.
sample fraction Is a number between 0 and 1 indicating what fraction
of the dataset to use in the automatic parameter configuration algorithm. Running the algorithm on the full dataset gives the most
accurate results, but for very large datasets can take longer than desired. In such case using just a fraction of the data helps speeding up
this algorithm while still giving good approximations of the optimum
parameters.
SavedIndexParams This object type is used for loading a previously
saved index from the disk.
struct SavedIndexParams : public IndexParams
{
SavedIndexParams( std::string filename );
};

filename The filename in which the index was saved.
3.1.2

flann::Index::buildIndex

Builds the nearest neighbor search index. There are two versions of the buildIndex
method, one that uses the points provided as argument and one that uses the
points provided to the constructor when the object was constructed.
void buildIndex();
void buildIndex(const Matrix& points);

11

3.1.3

flann::Index::addPoints

The addPoints method can be used to incrementally add points to the index
after the index was build. To avoid the index getting unbalanced, the addPoints
method has the option of rebuilding the index after a large number of points
have been added. The rebuild threshold parameter controls when the index
is rebuild, by default this is done when it doubles in size (rebuild threshold
= 2).
void addPoints(const Matrix& points, float rebuild_threshold = 2);

3.1.4

flann::Index::removePoint

The removePoint method removes one point with the specified point id from
the index. The indices (of the remaining points) returned by the nearest neighbor operations do not change when points are removed from the index.
void removePoint(size_t point_id);

3.1.5

flann::Index::getPoint

The getPoint method returns a pointer to the data point with the specified
point id.
ElementType* getPoint(size_t point_id);

3.1.6

flann::Index::knnSearch

Performs a K-nearest neighbor search for a set of query points. There are two
signatures for this method, one that takes pre-allocated flann::Matrix objects
for returning the indices of and distances to the neighbors found, and one that
takes std::vector that will re resized automatically as needed.
int Index::knnSearch(const Matrix& queries,
Matrix& indices,
Matrix& dists,
size_t knn,
const SearchParams& params);
int Index::knnSearch(const Matrix& queries,
std::vector< std::vector >& indices,
std::vector >& dists,
size_t knn,
const SearchParams& params);

queries Matrix containing the query points. Size of matrix is (num queries ×
dimentionality)

12

indices Matrix that will contain the indices of the K-nearest neighbors found
(size should be at least num queries × knn for the pre-allocated version).
dists Matrix that will contain the distances to the K-nearest neighbors found
(size should be at least num queries × knn for the pre-allocated version).
knn Number of nearest neighbors to search for.
params Search parameters. Structure containing parameters used during search.
SearchParameters
struct SearchParams
{
SearchParams(int checks = 32,
float eps = 0,
bool sorted = true);
int checks;
float eps;
bool sorted;
int max_neighbors;
tri_type use_heap;
int cores;
bool matrices_in_gpu_ram;
};

checks specifies the maximum leafs to visit when searching for neighbours. A higher value for this parameter would give better search
precision, but also take more time. For all leafs to be checked use
the value CHECKS UNLIMITED. If automatic configuration was used
when the index was created, the number of checks required to achieve
the specified precision was also computed, to use that value specify
CHECKS AUTOTUNED.
eps Search for eps-approximate neighbors (only used by KDTreeSingleIndex and KDTreeCuda3dIndex).
sorted Used only by radius search, specifies if the neighbors returned
should be sorted by distance.
max neighbours Specifies the maximum number of neighbors radius
search should return (default: -1 = unlimited). Only used for radius search.
use heap Use a heap data structure to manage the list of neighbors internally (possible values: FLANN False, FLANN True, FLANN Undefined).
A heap is more efficient for a large number of neighbors and less efficient for a small number of neighbors. Default value is FLANN Undefined,
which lets the code choose the best option depending on the number
of neighbors requested. Only used for KNN search.
cores How many cores to assign to the search (specify 0 for automatic
core selection).
matrices in gpu ram for GPU search indicates if matrices are already
in GPU ram.
13

3.1.7

flann::Index::radiusSearch

Performs a radius nearest neighbor search for a set of query points. There are
two signatures for this method, one that takes pre-allocated flann::Matrix
objects for returning the indices of and distances to the neighbors found, and
one that takes std::vector that will resized automatically as
needed.
int Index::radiusSearch(const Matrix& queries,
Matrix& indices,
Matrix& dists,
float radius,
const SearchParams& params);
int Index::radiusSearch(const Matrix& queries,
std::vector< std::vector >& indices,
std::vector >& dists,
float radius,
const SearchParams& params);

queries The query point. Size of matrix is (num queries × dimentionality).
indices Matrix that will contain the indices of the K-nearest neighbors found.
For the pre-allocated version, only as many neighbors are returned as
many columns in this matrix. If fewer neighbors are found than columns
in this matrix, the element after that last index returned is -1. In case
of the std::vector version, the rows will be resized as needed to fit all the
neighbors to be returned, except if the “max neighbors” search parameter
is set.
dists Matrix that will contain the distances to the K-nearest neighbors found.
The same number of values are returned here as for the indices matrix.
radius The search radius
params Search parameters
The method returns the number of nearest neighbors found.
3.1.8

flann::Index::save

Saves the index to a file.
void Index::save(std::string filename);

filename The file to save the index to

14

3.1.9

flann::hierarchicalClustering

Clusters the given points by constructing a hierarchical k-means tree and choosing a cut in the tree that minimizes the clusters’ variance.
template 
int hierarchicalClustering(const Matrix& features,
Matrix& centers,
const KMeansIndexParams& params,
Distance d = Distance())

features The points to be clustered
centers The centers of the clusters obtained. The number of rows in this
matrix represents the number of clusters desired. However, because of
the way the cut in the hierarchical tree is choosen, the number of clusters
computed will be the highest number of the form (branching − 1) ∗ k + 1
that’s lower than the number of clusters desired, where branching is the
tree’s branching factor (see description of the KMeansIndexParams).
params Parameters used in the construction of the hierarchical k-means tree
The function returns the number of clusters computed.
3.1.10

flann::KdTreeCuda3dIndex

FLANN provides a CUDA implementation of the kd-tree build and search algorithms to improve the build and query speed for large 3d data sets. This section
will provide all the necessary information to use the KdTreeCuda3dIndex index
type.
Building: If CMake detects a CUDA install during the build (see section
2), a library libflann cuda.so will be built.
Basic Usage: To be able to use the new index type, you have to include
the FLANN header this way:
#define FLANN_USE_CUDA
#include 

If you define the symbol FLANN USE CUDA before including the FLANN header,
you will have to link libflann cuda.so or libflann cuda s.a with your project.
However, you will not have to compile your source code with nvcc, except if
you use other CUDA code, of course.
You can then create your index by using the KDTreeCuda3dIndexParams to
create the index. The index will take care of copying all the data from and to
the GPU for you, both for index creation and search.
A word of caution: A general guideline for deciding whether to use the
CUDA kd tree or a (multi-threaded) CPU implementation is hard to give, since
it depends on the combination of CPU and GPU in each system and the data
sets. For example, on a system with a Phenom II 955 CPU and a Geforce GTX
15

260 GPU, the maximum search speedup on a synthetic (random) data set is a
factor of about 8-9 vs the single-core CPU search, and starts to be reached at
about 100k search and query points. (This includes transfer times.) Build time
does not profit as much from the GPU acceleration; here the benefit is about
2x at 200k points, but this largely depends on the data set. The only way to
know which implementation is best suited is to try it.
Advanced Usage: In some cases, you might already have your data in a
buffer on the GPU. In this case, you can re-use these buffers instead of copying
the buffers back to system RAM for index creation and search. The way to do
this is to pass GPU pointers via the flann::Matrix inputs and tell the index
via the index and search params to treat the pointers as GPU pointers.
thrust::device_vector
// ... fill vector here...

points_vector( n_points );

float* gpu_pointer = (float*)thrust::raw_pointer_cast(&points_vector[0]);
flann::Matrix matrix_gpu(gpu_pointer,n_points,3, 4);
flann::KDTreeCuda3dIndexParams params;
params["input_is_gpu_float4"]=true;
flann::Index > flannindex( matrix_gpu, params
flannindex.buildIndex();

);

First, a GPU buffer of float4 values is created and filled with points.

2

Then, a GPU pointer to the buffer is stored in a flann matrix with 3 columns
and a stride of 4 (since the last element in the float4 is unused).
Last, the index is created. The input is gpu float4 flag tells the index to
treat the matrix as a gpu buffer.
Similarly, you can specify GPU buffers for the search routines that return the
result via flann matrices (but not for those that return them via std::vectors).
To do this, the pointers in the index and dist matrices have to point to GPU
buffers and the cols value has to be set to 3 (as we are dealing with 3d points).
Here, any value for stride can be used.
flann::Matrix indices_gpu(gpu_pointer_indices,n_points, knn, stride);
flann::Matrix dists_gpu(gpu_pointer_dists,n_points, knn, stride);
flann::SearchParams params;
params.matrices_in_gpu_ram = true;
flannindex.knnSearch( queries_gpu ,indices_gpu,dists_gpu,knn, params);

Note that you cannot mix matrices in system and CPU ram here!
2 For index creation, only float4 points are supported, float3 or structure-of-array (SOA)
representations are currently not supported since float4 proved to be best in terms of access
speed for tree creation and search.

16

Search Parameters: The search routines support three parameters:
• eps - used for approximate knn search. The maximum possible error is
e = dbest ∗ eps, i.e. the distance of the returned neighbor is at maximum
eps times larget than the distance to the real best neighbor.
• use heap - used in knn and radius search. If set to true, a heap structure
will be used in the search to keep track of the distance to the farthest
neighbor. Beneficial with about k > 64 elements.
• sorted - if set to true, the results of the radius and knn searches will be
sorted in ascending order by their distance to the query point.
• matrices in gpu ram - set to true to indicate that all (input and output)
matrices are located in GPU RAM.

3.2

Using FLANN from C

FLANN can be used in C programs through the C bindings provided with the
library. Because there is no template support in C, there are bindings provided
for the following data types: unsigned char, int, float and double. For each
of the functions below there is a corresponding version for each of the for data
types, for example for the function:
flan_index_t flann_build_index(float* dataset, int rows, int cols, float* speedup,
struct FLANNParameters* flann_params);

there are also the following versions:
flan_index_t flann_build_index_byte(unsigned char* dataset,
int rows, int cols, float* speedup,
struct FLANNParameters* flann_params);
flan_index_t flann_build_index_int(int* dataset,
int rows, int cols, float* speedup,
struct FLANNParameters* flann_params);
flan_index_t flann_build_index_float(float* dataset,
int rows, int cols, float* speedup,
struct FLANNParameters* flann_params);
flan_index_t flann_build_index_double(double* dataset,
int rows, int cols, float* speedup,
struct FLANNParameters* flann_params);

3.2.1

flann build index()

flan_index_t flann_build_index(float* dataset,
int rows,
int cols,
float* speedup,
struct FLANNParameters* flann_params);

17

This function builds an index and return a reference to it. The arguments
expected by this function are as follows:
dataset, rows and cols - are used to specify the input dataset of points:
dataset is a pointer to a rows × cols matrix stored in row-major order
(one feature on each row)
speedup - is used to return the approximate speedup over linear search achieved
when using the automatic index and parameter configuration (see section
3.3.1)
flann params - is a structure containing the parameters passed to the function.
This structure is defined as follows:
struct FLANNParameters {
enum flann_algorithm_t algorithm;

/* the algorithm to use */

/* search parameters */
int checks;
/* how many leafs (features) to check in one search */
float cb_index;
/* cluster boundary index. Used when searching the
kmeans tree */
/* kdtree index parameters */
int trees;
/* number of randomized trees to use (for kdtree) */
/* kmeans index parameters */
int branching;
/* branching factor (for kmeans tree) */
int iterations;
/* max iterations to perform in one kmeans cluetering
(kmeans tree) */
enum flann_centers_init_t centers_init; /* algorithm used for picking the initial
cluster centers for kmeans tree */
/* autotuned index parameters
float target_precision;
/*
float build_weight;
/*
float memory_weight;
/*
float sample_fraction;
/*

*/
precision desired (used for autotuning, -1 otherwise) */
build tree time weighting factor */
index memory weigthing factor */
what fraction of the dataset to use for autotuning */

/* LSH parameters */
unsigned int table_number_; /** The number of hash tables to use */
unsigned int key_size_;
/** The length of the key in the hash tables */
unsigned int multi_probe_level_; /** Number of levels to use in multi-probe LSH, 0 for standard LSH */
/* other parameters */
enum flann_log_level_t log_level;
long random_seed;
};

/* determines the verbosity of each flann function */
/* random seed to use */

The algorithm and centers init fields can take the following values:
enum flann_algorithm_t {
FLANN_INDEX_LINEAR = 0,
FLANN_INDEX_KDTREE = 1,
FLANN_INDEX_KMEANS = 2,
FLANN_INDEX_COMPOSITE = 3,
FLANN_INDEX_KDTREE_SINGLE = 3,

18

FLANN_INDEX_SAVED = 254,
FLANN_INDEX_AUTOTUNED = 255
};
enum flann_centers_init_t {
FLANN_CENTERS_RANDOM = 0,
FLANN_CENTERS_GONZALES = 1,
FLANN_CENTERS_KMEANSPP = 2
};

The algorithm field is used to manually select the type of index used.
The centers init field specifies how to choose the inital cluster centers when performing the hierarchical k-means clustering (in case the
algorithm used is k-means): FLANN CENTERS RANDOM chooses the initial
centers randomly, FLANN CENTERS GONZALES chooses the initial centers
to be spaced apart from each other by using Gonzales’ algorithm and
FLANN CENTERS KMEANSPP chooses the initial centers using the algorithm
proposed in [AV07].
The fields: checks, cb index, trees, branching, iterations, target precision,
build weight, memory weight and sample fraction have the same meaning as described in 3.1.1.
The random seed field contains the random seed useed to initialize the
random number generator.
The field log level controls the verbosity of the messages generated by
the FLANN library functions. It can take the following values:
enum flann_log_level_t {
FLANN_LOG_NONE = 0,
FLANN_LOG_FATAL = 1,
FLANN_LOG_ERROR = 2,
FLANN_LOG_WARN = 3,
FLANN_LOG_INFO = 4
};

3.2.2

flann find nearest neighbors index()

int flann_find_nearest_neighbors_index(FLANN_INDEX index_id,
float* testset,
int trows,
int* indices,
float* dists,
int nn,
int checks,
struct FLANNParameters* flann_params);

This function searches for the nearest neighbors of the testset points using
an index already build and referenced by index id. The testset is a matrix stored in row-major format with trows rows and the same number of
columns as the dimensionality of the points used to build the index. The function computes nn nearest neighbors for each point in the testset and stores
19

them in the indices matrix (which is a trows × nn matrix stored in row-major
format). The memory for the result matrix must be allocated before the
flann find nearest neighbors index() function is called. The distances to
the nearest neighbors found are stored in the dists matrix. The checks parameter specifies how many tree traversals should be performed during the search.
3.2.3

flann find nearest neighbors()

int flann_find_nearest_neighbors(float* dataset,
int rows,
int cols,
float* testset,
int trows,
int* indices,
float* dists,
int nn,
struct FLANNParameters* flann_params);

This function is similar to the flann find nearest neighbors index() function, but instread of using a previously constructed index, it constructs the
index, does the nearest neighbor search and deletes the index in one step.
3.2.4

flann radius search()

int flann_radius_search(FLANN_INDEX index_ptr,
float* query,
/* query point */
int* indices, /* array for storing the indices */
float* dists, /* similar, but for storing distances */
int max_nn, /* size of arrays indices and dists */
float radius, /* search radius (squared radius for euclidian metric) */
int checks, /* number of features to check, sets the level
of approximation */
FLANNParameters* flann_params);

This function performs a radius search to single query point. The indices of
the neighbors found and the distances to them are stored in the indices and
dists arrays. The max nn parameter sets the limit of the neighbors that will be
returned (the size of the indices and dists arrays must be at least max nn).
3.2.5

flann save index()

int flann_save_index(flann_index_t index_id,
char* filename);

This function saves an index to a file. The dataset for which the index was
built is not saved with the index.

20

3.2.6

flann load index()

flann_index_t flann_load_index(char* filename,
float* dataset,
int rows,
int cols);

This function loads a previously saved index from a file. Since the dataset
is not saved with the index, it must be provided to this function.
3.2.7

flann free index()

int flann_free_index(FLANN_INDEX index_id,
struct FLANNParameters* flann_params);

This function deletes a previously constructed index and frees all the memory
used by it.
3.2.8

flann set distance type

This function chooses the distance function to use when computing distances
between data points.
void flann_set_distance_type(enum flann_distance_t distance_type, int order);

distance type The distance type to use. Possible values are
enum flann_distance_t {
FLANN_DIST_EUCLIDEAN = 1, // squared euclidean distance
FLANN_DIST_MANHATTAN = 2,
FLANN_DIST_MINKOWSKI = 3,
FLANN_DIST_HIST_INTERSECT = 5,
FlANN_DIST_HELLINGER = 6,
FLANN_DIST_CHI_SQUARE = 7,
// chi-square
FLANN_DIST_KULLBACK_LEIBLER = 8,
// kullback-leibler divergence
};

order Used in for the FLANN DIST MINKOWSKI distance type, to choose the order
of the Minkowski distance.
3.2.9

flann compute cluster centers()

Performs hierarchical clustering of a set of points (see 3.1.9).
int flann_compute_cluster_centers(float* dataset,
int rows,
int cols,
int clusters,
float* result,
struct FLANNParameters* flann_params);

See section 1.1 for an example of how to use the C/C++ bindings.
21

3.3

Using FLANN from MATLAB

The FLANN library can be used from MATLAB through the following wrapper
functions: flann build index, flann search, flann save index, flann load index,
flann set distance type and flann free index. The flann build index
function creates a search index from the dataset points, flann search uses this
index to perform nearest-neighbor searches, flann save index and flann load index
can be used to save and load an index to/from disk, flann set distance type is
used to set the distance type to be used when building an index and flann free index
deletes the index and releases the memory it uses.
The following sections describe in more detail the FLANN matlab wrapper
functions and show examples of how they may be used.
3.3.1

flann build index

This function creates a search index from the initial dataset of points, index
used later for fast nearest-neighbor searches in the dataset.
[index, parameters, speedup] = flann_build_index(dataset, build_params);

The arguments passed to the flann build index function have the following
meaning:
dataset is a d×n matrix containing n d-dimensional points, stored in a columnmajor order (one feature on each column)
build params - is a MATLAB structure containing the parameters passed to
the function.
The build params is used to specify the type of index to be built and the
index parameters. These have a big impact on the performance of the new search
index (nearest-neighbor search time) and on the time and memory required
to build the index. The optimum parameter values depend on the dataset
characteristics (number of dimensions, distribution of points in the dataset)
and on the application domain (desired precision for the approximate nearest
neighbor searches). The build params argument is a structure that contains
one or more of the following fields:
algorithm - the algorithm to use for building the index. The possible values are: ’linear’, ’kdtree’, ’kmeans’, ’composite’ or ’autotuned’.
The ’linear’ option does not create any index, it uses brute-force search
in the original dataset points, ’kdtree’ creates one or more randomized kd-trees, ’kmeans’ creates a hierarchical kmeans clustering tree,
’composite’ is a mix of both kdtree and kmeans trees and the ’autotuned’
automatically determines the best index and optimum parameters using
a cross-validation technique.
Autotuned index: in case the algorithm field is ’autotuned’, the following
fields should also be present:
22

target precision - is a number between 0 and 1 specifying the percentage of
the approximate nearest-neighbor searches that return the exact nearestneighbor. Using a higher value for this parameter gives more accurate
results, but the search takes longer. The optimum value usually depends
on the application.
build weight - specifies the importance of the index build time raported to
the nearest-neighbor search time. In some applications it’s acceptable for
the index build step to take a long time if the subsequent searches in the
index can be performed very fast. In other applications it’s required that
the index be build as fast as possible even if that leads to slightly longer
search times. (Default value: 0.01)
memory weight - is used to specify the tradeoff between time (index build time
and search time) and memory used by the index. A value less than 1 gives
more importance to the time spent and a value greater than 1 gives more
importance to the memory usage.
sample fraction - is a number between 0 and 1 indicating what fraction of
the dataset to use in the automatic parameter configuration algorithm.
Running the algorithm on the full dataset gives the most accurate results,
but for very large datasets can take longer than desired. In such case,
using just a fraction of the data helps speeding up this algorithm, while
still giving good approximations of the optimum parameters.
Randomized kd-trees index: in case the algorithm field is ’kdtree’, the
following fields should also be present:
trees - the number of randomized kd-trees to create.
Hierarchical k-means index: in case the algorithm type is ’means’, the
following fields should also be present:
branching - the branching factor to use for the hierarchical kmeans tree creation. While kdtree is always a binary tree, each node in the kmeans tree
may have several branches depending on the value of this parameter.
iterations - the maximum number of iterations to use in the kmeans clustering
stage when building the kmeans tree. A value of -1 used here means that
the kmeans clustering should be performed until convergence.
centers init - the algorithm to use for selecting the initial centers when performing a kmeans clustering step. The possible values are ’random’ (picks
the initial cluster centers randomly), ’gonzales’ (picks the initial centers
using the Gonzales algorithm) and ’kmeanspp’ (picks the initial centers
using the algorithm suggested in [AV07]). If this parameters is omitted,
the default value is ’random’.
23

cb index - this parameter (cluster boundary index) influences the way exploration is performed in the hierarchical kmeans tree. When cb index is
zero the next kmeans domain to be explored is choosen to be the one with
the closest center. A value greater then zero also takes into account the
size of the domain.
Composite index: in case the algorithm type is ’composite’, the fields from
both randomized kd-tree index and hierarchical k-means index should be
present.
The flann build index function returns the newly created index, the parameters
used for creating the index and, if automatic configuration was used, an estimation of the speedup over linear search that is achieved when searching the
index. Since the parameter estimation step is costly, it is possible to save the
computed parameters and reuse them the next time an index is created from
similar data points (coming from the same distribution).
3.3.2

flann search

This function performs nearest-neighbor searches using the index already created:
[result, dists] = flann_search(index, testset, k, parameters);

The arguments required by this function are:
index - the index returned by the flann build index function
testset - a d × m matrix containing m test points whose k-nearest-neighbors
need to be found
k - the number of nearest neighbors to be returned for each point from testset
parameters - structure containing the search parameters. Currently it has
only one member, parameters.checks, denoting the number of times the
tree(s) in the index should be recursively traversed. A higher value for this
parameter would give better search precision, but also take more time. If
automatic configuration was used when the index was created, the number
of checks required to achieve the specified precision is also computed. In
such case, the parameters structure returned by the flann build index
function can be passed directly to the flann search function.
The function returns two matrices, each of size k ×m. The first one contains,
in which each column, the indexes (in the dataset matrix) of the k nearest
neighbors of the corresponding point from testset, while the second one contains
the corresponding distances. The second matrix can be omitted when making
the call if the distances to the nearest neighbors are not needed.

24

For the case where a single search will be performed with each index, the
flann search function accepts the dataset instead of the index as first argument, in which case the index is created searched and then deleted in one step.
In this case the parameters structure passed to the flann search function must
also contain the fields of the build params structure that would normally be
passed to the flann build index function if the index was build separately.
[result, dists] = flann_search(dataset, testset, k, parameters);
3.3.3

flann save index

This function saves an index to a file so that it can be reused at a later time
without the need to recompute it. Only the index will be saved to the file, not
also the data points for which the index was created, so for the index to be
reused the data points must be saved separately.
flann_save_index(index, filename)

The argumenst required by this function are:
index - the index to be saved, created by flann build index
filename - the name of the file in which to save the index
3.3.4

flann load index

This function loads a previously saved index from a file. It needs to be passed
as a second parameter the dataset for which the index was created, as this is
not saved together with the index.
index = flann_load_index(filename, dataset)

The argumenst required by this function are:
filename - the file from which to load the index
dataset - the dataset for which the index was created
This function returns the index object.
3.3.5

flann set distance type

This function chooses the distance function to use when computing distances
between data points.
flann_set_distance_type(type, order)

The argumenst required by this function are:

25

type - the distance type to use. Possible values are: ’euclidean’, ’manhattan’,
’minkowski’, ’max dist’ (L inf inity - distance type is not valid for kdtree index type since it’s not dimensionwise additive), ’hik’ (histogram
intersection kernel), ’hellinger’,’cs’ (chi-square) and ’kl’ (KullbackLeibler).
order - only used if distance type is ’minkowski’ and represents the order of
the minkowski distance.
3.3.6

flann free index

This function must be called to delete an index and release all the memory used
by it:
flann_free_index(index);

3.3.7

Examples

Let’s look at a few examples showing how the functions described above are
used:
3.3.8

Example 1:

In this example the index is constructed using automatic parameter estimation,
requesting 70% as desired precision and using the default values for the build
time and memory usage factors. The index is then used to search for the nearestneighbors of the points in the testset matrix and finally the index is deleted.
dataset = single(rand(128,10000));
testset = single(rand(128,1000));
build_params.algorithm = ’autotuned’;
build_params.target_precision = 0.7;
build_params.build_weight = 0.01;
build_params.memory_weight = 0;
[index, parameters] = flann_build_index(dataset, build_params);
result = flann_search(index,testset,5,parameters);
flann_free_index(index);

3.3.9

Example 2:

In this example the index constructed with the parameters specified manually.
dataset = single(rand(128,10000));

26

testset = single(rand(128,1000));
index = flann_build_index(dataset,struct(’algorithm’,’kdtree’,’trees’,8));
result = flann_search(index,testset,5,struct(’checks’,128));
flann_free_index(index);

3.3.10

Example 3:

In this example the index creation, searching and deletion are all performed in
one step:
dataset = single(rand(128,10000));
testset = single(rand(128,1000));
[result,dists] = flann_search(dataset,testset,5,struct(’checks’,128,’algorithm’,...
’kmeans’,’branching’,64,’iterations’,5));

3.4

Using FLANN from python

The python bindings are automatically installed on the system with the FLANN
library if the option BUILD PYTHON BINDINGS is enabled. You may need to set
the PYTHONPATH to the location of the bindings if you installed FLANN in a
non-standard location. The python bindings also require the numpy package to
be installed.
To use the python FLANN bindings the package pyflann must be imported
(see the python example in section 1.1). This package contains a class called
FLANN that handles the nearest-neighbor search operations. This class containg the following methods:
def build index(self, dataset, **kwargs) :
This method builds and internally stores an index to be used for future
nearest neighbor matchings. It erases any previously stored index, so in
order to work with multiple indexes, multiple instances of the FLANN
class must be used. The dataset argument must be a 2D numpy array
or a matrix, stored in a row-major order (a feature on each row of the
matrix). The rest of the arguments that can be passed to the method are
the same as those used in the build params structure from section 3.3.1.
Similar to the MATLAB version, the index can be created using manually
specified parameters or the parameters can be automatically computed
(by specifying the target precision, build weight and memory weight arguments).
The method returns a dictionary containing the parameters used to construct the index. In case automatic parameter selection is used, the dictionary will also contain the number of checks required to achieve the desired
27

target precision and an estimation of the speedup over linear search that
the library will provide.
def nn index(self, testset, num neighbors = 1, **kwargs) :
This method searches for the num neighbors nearest neighbors of each
point in testset using the index computed by build index. Additionally,
a parameter called checks, denoting the number of times the index tree(s)
should be recursivelly searched, must be given.
Example:
from pyflann import *
from numpy import *
from numpy.random import *
dataset = rand(10000, 128)
testset = rand(1000, 128)
flann = FLANN()
params = flann.build_index(dataset, algorithm="autotuned", target_precision=0.9, log_level = "info");
print params
result, dists = flann.nn_index(testset,5, checks=params["checks"]);

def nn(self, dataset, testset, num neighbors = 1, **kwargs) :
This method builds the index, performs the nearest neighbor search and
deleted the index, all in one step.
def save index(self, filename) :
This method saves the index to a file. The dataset form which the index
was build is not saved.
def load index(self, filename, pts) :
Load the index from a file. The dataset for which the index was build
must also be provided since is not saved with the index.
def set distance type(distance type, order = 0) :
This function (part of the pyflann module) sets the distance type to be
used. See 3.3.5 for possible values of the distance type.
See section 1.1 for an example of how to use the Python bindings.

References
[AV07] D. Arthur and S. Vassilvitskii. k-means++: the advantages of careful
seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027–1035. Society for Industrial
and Applied Mathematics Philadelphia, PA, USA, 2007.

28



---

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.5
Linearized                      : No
Page Count                      : 29
Page Mode                       : UseOutlines
Author                          : 
Title                           : 
Subject                         : 
Creator                         : LaTeX with hyperref package
Producer                        : pdfTeX-1.40.13
Create Date                     : 2013:01:24 09:53:02-08:00
Modify Date                     : 2013:01:24 09:53:02-08:00
Trapped                         : False
PTEX Fullbanner                 : This is pdfTeX, Version 3.1415926-2.4-1.40.13 (TeX Live 2012/Debian) kpathsea version 6.1.0

EXIF Metadata provided by EXIF.tools