/***********************************************************************
* Software License Agreement (BSD License)
*
* Copyright 2008-2011 Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
* Copyright 2008-2011 David G. Lowe (lowe@cs.ubc.ca). All rights reserved.
*
* THE BSD LICENSE
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*************************************************************************/
#ifndef OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_
#define OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_
#include <algorithm>
#include <map>
#include <cassert>
#include <limits>
#include <cmath>
#include "general.h"
#include "nn_index.h"
#include "dist.h"
#include "matrix.h"
#include "result_set.h"
#include "heap.h"
#include "allocator.h"
#include "random.h"
#include "saving.h"
namespace cvflann
{
struct HierarchicalClusteringIndexParams : public IndexParams
{
HierarchicalClusteringIndexParams(int branching = 32,
flann_centers_init_t centers_init = FLANN_CENTERS_RANDOM,
int trees = 4, int leaf_size = 100)
{
(*this)["algorithm"] = FLANN_INDEX_HIERARCHICAL;
// The branching factor used in the hierarchical clustering
(*this)["branching"] = branching;
// Algorithm used for picking the initial cluster centers
(*this)["centers_init"] = centers_init;
// number of parallel trees to build
(*this)["trees"] = trees;
// maximum leaf size
(*this)["leaf_size"] = leaf_size;
}
};
/**
* Hierarchical index
*
* Contains a tree constructed through a hierarchical clustering
* and other information for indexing a set of points for nearest-neighbour matching.
*/
template <typename Distance>
class HierarchicalClusteringIndex : public NNIndex<Distance>
{
public:
typedef typename Distance::ElementType ElementType;
typedef typename Distance::ResultType DistanceType;
private:
typedef void (HierarchicalClusteringIndex::* centersAlgFunction)(int, int*, int, int*, int&);
/**
* The function used for choosing the cluster centers.
*/
centersAlgFunction chooseCenters;
/**
* Chooses the initial centers in the k-means clustering in a random manner.
*
* Params:
* k = number of centers
* vecs = the dataset of points
* indices = indices in the dataset
* indices_length = length of indices vector
*
*/
void chooseCentersRandom(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
{
UniqueRandom r(indices_length);
int index;
for (index=0; index<k; ++index) {
bool duplicate = true;
int rnd;
while (duplicate) {
duplicate = false;
rnd = r.next();
if (rnd<0) {
centers_length = index;
return;
}
centers[index] = dsindices[rnd];
for (int j=0; j<index; ++j) {
DistanceType sq = distance(dataset[centers[index]], dataset[centers[j]], dataset.cols);
if (sq<1e-16) {
duplicate = true;
}
}
}
}
centers_length = index;
}
/**
* Chooses the initial centers in the k-means using Gonzales' algorithm
* so that the centers are spaced apart from each other.
*
* Params:
* k = number of centers
* vecs = the dataset of points
* indices = indices in the dataset
* Returns:
*/
void chooseCentersGonzales(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
{
int n = indices_length;
int rnd = rand_int(n);
assert(rnd >=0 && rnd < n);
centers[0] = dsindices[rnd];
int index;
for (index=1; index<k; ++index) {
int best_index = -1;
DistanceType best_val = 0;
for (int j=0; j<n; ++j) {
DistanceType dist = distance(dataset[centers[0]],dataset[dsindices[j]],dataset.cols);
for (int i=1; i<index; ++i) {
DistanceType tmp_dist = distance(dataset[centers[i]],dataset[dsindices[j]],dataset.cols);
if (tmp_dist<dist) {
dist = tmp_dist;
}
}
if (dist>best_val) {
best_val = dist;
best_index = j;
}
}
if (best_index!=-1) {
centers[index] = dsindices[best_index];
}
else {
break;
}
}
centers_length = index;
}
/**
* Chooses the initial centers in the k-means using the algorithm
* proposed in the KMeans++ paper:
* Arthur, David; Vassilvitskii, Sergei - k-means++: The Advantages of Careful Seeding
*
* Implementation of this function was converted from the one provided in Arthur's code.
*
* Params:
* k = number of centers
* vecs = the dataset of points
* indices = indices in the dataset
* Returns:
*/
void chooseCentersKMeanspp(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
{
int n = indices_length;
double currentPot = 0;
DistanceType* closestDistSq = new DistanceType[n];
// Choose one random center and set the closestDistSq values
int index = rand_int(n);
assert(index >=0 && index < n);
centers[0] = dsindices[index];
// Computing distance^2 will have the advantage of even higher probability further to pick new centers
// far from previous centers (and this complies to "k-means++: the advantages of careful seeding" article)
for (int i = 0; i < n; i++) {
closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
closestDistSq[i] = ensureSquareDistance<Distance>( closestDistSq[i] );
currentPot += closestDistSq[i];
}
const int numLocalTries = 1;
// Choose each center
int centerCount;
for (centerCount = 1; centerCount < k; centerCount++) {
// Repeat several trials
double bestNewPot = -1;
int bestNewIndex = 0;
for (int localTrial = 0; localTrial < numLocalTries; localTrial++) {
// Choose our center - have to be slightly careful to return a valid answer even accounting
// for possible rounding errors
double randVal = rand_double(currentPot);
for (index = 0; index < n-1; index++) {
if (randVal <= closestDistSq[index]) break;
else randVal -= closestDistSq[index];
}
// Compute the new potential
double newPot = 0;
for (int i = 0; i < n; i++) {
DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
newPot += std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
}
// Store the best result
if ((bestNewPot < 0)||(newPot < bestNewPot)) {
bestNewPot = newPot;
bestNewIndex = index;
}
}
// Add the appropriate center
centers[centerCount] = dsindices[bestNewIndex];
currentPot = bestNewPot;
for (int i = 0; i < n; i++) {
DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols);
closestDistSq[i] = std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
}
}
centers_length = centerCount;
delete[] closestDistSq;
}
/**
* Chooses the initial centers in a way inspired by Gonzales (by Pierre-Emmanuel Viel):
* select the first point of the list as a candidate, then parse the points list. If another
* point is further than current candidate from the other centers, test if it is a good center
* of a local aggregation. If it is, replace current candidate by this point. And so on...
*
* Used with KMeansIndex that computes centers coordinates by averaging positions of clusters points,
* this doesn't make a real difference with previous methods. But used with HierarchicalClusteringIndex
* class that pick centers among existing points instead of computing the barycenters, there is a real
* improvement.
*
* Params:
* k = number of centers
* vecs = the dataset of points
* indices = indices in the dataset
* Returns:
*/
void GroupWiseCenterChooser(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
{
const float kSpeedUpFactor = 1.3f;
int n = indices_length;
DistanceType* closestDistSq = new DistanceType[n];
// Choose one random center and set the closestDistSq values
int index = rand_int(n);
assert(index >=0 && index < n);
centers[0] = dsindices[index];
for (int i = 0; i < n; i++) {
closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
}
// Choose each center
int centerCount;
for (centerCount = 1; centerCount < k; centerCount++) {
// Repeat several trials
double bestNewPot = -1;
int bestNewIndex = 0;
DistanceType furthest = 0;
for (index = 0; index < n; index++) {
// We will test only the potential of the points further than current candidate
if( closestDistSq[index] > kSpeedUpFactor * (float)furthest ) {
// Compute the new potential
double newPot = 0;
for (int i = 0; i < n; i++) {
newPot += std::min( distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols)
, closestDistSq[i] );
}
// Store the best result
if ((bestNewPot < 0)||(newPot <= bestNewPot)) {
bestNewPot = newPot;
bestNewIndex = index;
furthest = closestDistSq[index];
}
}
}
// Add the appropriate center
centers[centerCount] = dsindices[bestNewIndex];
for (int i = 0; i < n; i++) {
closestDistSq[i] = std::min( distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols)
, closestDistSq[i] );
}
}
centers_length = centerCount;
delete[] closestDistSq;
}
public:
/**
* Index constructor
*
* Params:
* inputData = dataset with the input features
* params = parameters passed to the hierarchical k-means algorithm
*/
HierarchicalClusteringIndex(const Matrix<ElementType>& inputData, const IndexParams& index_params = HierarchicalClusteringIndexParams(),
Distance d = Distance())
: dataset(inputData), params(index_params), root(NULL), indices(NULL), distance(d)
{
memoryCounter = 0;
size_ = dataset.rows;
veclen_ = dataset.cols;
branching_ = get_param(params,"branching",32);
centers_init_ = get_param(params,"centers_init", FLANN_CENTERS_RANDOM);
trees_ = get_param(params,"trees",4);
leaf_size_ = get_param(params,"leaf_size",100);
if (centers_init_==FLANN_CENTERS_RANDOM) {
chooseCenters = &HierarchicalClusteringIndex::chooseCentersRandom;
}
else if (centers_init_==FLANN_CENTERS_GONZALES) {
chooseCenters = &HierarchicalClusteringIndex::chooseCentersGonzales;
}
else if (centers_init_==FLANN_CENTERS_KMEANSPP) {
chooseCenters = &HierarchicalClusteringIndex::chooseCentersKMeanspp;
}
else if (centers_init_==FLANN_CENTERS_GROUPWISE) {
chooseCenters = &HierarchicalClusteringIndex::GroupWiseCenterChooser;
}
else {
throw FLANNException("Unknown algorithm for choosing initial centers.");
}
trees_ = get_param(params,"trees",4);
root = new NodePtr[trees_];
indices = new int*[trees_];
for (int i=0; i<trees_; ++i) {
root[i] = NULL;
indices[i] = NULL;
}
}
HierarchicalClusteringIndex(const HierarchicalClusteringIndex&);
HierarchicalClusteringIndex& operator=(const HierarchicalClusteringIndex&);
/**
* Index destructor.
*
* Release the memory used by the index.
*/
virtual ~HierarchicalClusteringIndex()
{
free_elements();
if (root!=NULL) {
delete[] root;
}
if (indices!=NULL) {
delete[] indices;
}
}
/**
* Release the inner elements of indices[]
*/
void free_elements()
{
if (indices!=NULL) {
for(int i=0; i<trees_; ++i) {
if (indices[i]!=NULL) {
delete[] indices[i];
indices[i] = NULL;
}
}
}
}
/**
* Returns size of index.
*/
size_t size() const
{
return size_;
}
/**
* Returns the length of an index feature.
*/
size_t veclen() const
{
return veclen_;
}
/**
* Computes the inde memory usage
* Returns: memory used by the index
*/
int usedMemory() const
{
return pool.usedMemory+pool.wastedMemory+memoryCounter;
}
/**
* Builds the index
*/
void buildIndex()
{
if (branching_<2) {
throw FLANNException("Branching factor must be at least 2");
}
free_elements();
for (int i=0; i<trees_; ++i) {
indices[i] = new int[size_];
for (size_t j=0; j<size_; ++j) {
indices[i][j] = (int)j;
}
root[i] = pool.allocate<Node>();
computeClustering(root[i], indices[i], (int)size_, branching_,0);
}
}
flann_algorithm_t getType() const
{
return FLANN_INDEX_HIERARCHICAL;
}
void saveIndex(FILE* stream)
{
save_value(stream, branching_);
save_value(stream, trees_);
save_value(stream, centers_init_);
save_value(stream, leaf_size_);
save_value(stream, memoryCounter);
for (int i=0; i<trees_; ++i) {
save_value(stream, *indices[i], size_);
save_tree(stream, root[i], i);
}
}
void loadIndex(FILE* stream)
{
free_elements();
if (root!=NULL) {
delete[] root;
}
if (indices!=NULL) {
delete[] indices;
}
load_value(stream, branching_);
load_value(stream, trees_);
load_value(stream, centers_init_);
load_value(stream, leaf_size_);
load_value(stream, memoryCounter);
indices = new int*[trees_];
root = new NodePtr[trees_];
for (int i=0; i<trees_; ++i) {
indices[i] = new int[size_];
load_value(stream, *indices[i], size_);
load_tree(stream, root[i], i);
}
params["algorithm"] = getType();
params["branching"] = branching_;
params["trees"] = trees_;
params["centers_init"] = centers_init_;
params["leaf_size"] = leaf_size_;
}
/**
* Find set of nearest neighbors to vec. Their indices are stored inside
* the result object.
*
* Params:
* result = the result object in which the indices of the nearest-neighbors are stored
* vec = the vector for which to search the nearest neighbors
* searchParams = parameters that influence the search algorithm (checks)
*/
void findNeighbors(ResultSet<DistanceType>& result, const ElementType* vec, const SearchParams& searchParams)
{
int maxChecks = get_param(searchParams,"checks",32);
// Priority queue storing intermediate branches in the best-bin-first search
Heap<BranchSt>* heap = new Heap<BranchSt>((int)size_);
std::vector<bool> checked(size_,false);
int checks = 0;
for (int i=0; i<trees_; ++i) {
findNN(root[i], result, vec, checks, maxChecks, heap, checked);
}
BranchSt branch;
while (heap->popMin(branch) && (checks<maxChecks || !result.full())) {
NodePtr node = branch.node;
findNN(node, result, vec, checks, maxChecks, heap, checked);
}
assert(result.full());
delete heap;
}
IndexParams getParameters() const
{
return params;
}
private:
/**
* Struture representing a node in the hierarchical k-means tree.
*/
struct Node
{
/**
* The cluster center index
*/
int pivot;
/**
* The cluster size (number of points in the cluster)
*/
int size;
/**
* Child nodes (only for non-terminal nodes)
*/
Node** childs;
/**
* Node points (only for terminal nodes)
*/
int* indices;
/**
* Level
*/
int level;
};
typedef Node* NodePtr;
/**
* Alias definition for a nicer syntax.
*/
typedef BranchStruct<NodePtr, DistanceType> BranchSt;
void save_tree(FILE* stream, NodePtr node, int num)
{
save_value(stream, *node);
if (node->childs==NULL) {
int indices_offset = (int)(node->indices - indices[num]);
save_value(stream, indices_offset);
}
else {
for(int i=0; i<branching_; ++i) {
save_tree(stream, node->childs[i], num);
}
}
}
void load_tree(FILE* stream, NodePtr& node, int num)
{
node = pool.allocate<Node>();
load_value(stream, *node);
if (node->childs==NULL) {
int indices_offset;
load_value(stream, indices_offset);
node->indices = indices[num] + indices_offset;
}
else {
node->childs = pool.allocate<NodePtr>(branching_);
for(int i=0; i<branching_; ++i) {
load_tree(stream, node->childs[i], num);
}
}
}
void computeLabels(int* dsindices, int indices_length, int* centers, int centers_length, int* labels, DistanceType& cost)
{
cost = 0;
for (int i=0; i<indices_length; ++i) {
ElementType* point = dataset[dsindices[i]];
DistanceType dist = distance(point, dataset[centers[0]], veclen_);
labels[i] = 0;
for (int j=1; j<centers_length; ++j) {
DistanceType new_dist = distance(point, dataset[centers[j]], veclen_);
if (dist>new_dist) {
labels[i] = j;
dist = new_dist;
}
}
cost += dist;
}
}
/**
* The method responsible with actually doing the recursive hierarchical
* clustering
*
* Params:
* node = the node to cluster
* indices = indices of the points belonging to the current node
* branching = the branching factor to use in the clustering
*
* TODO: for 1-sized clusters don't store a cluster center (it's the same as the single cluster point)
*/
void computeClustering(NodePtr node, int* dsindices, int indices_length, int branching, int level)
{
node->size = indices_length;
node->level = level;
if (indices_length < leaf_size_) { // leaf node
node->indices = dsindices;
std::sort(node->indices,node->indices+indices_length);
node->childs = NULL;
return;
}
std::vector<int> centers(branching);
std::vector<int> labels(indices_length);
int centers_length;
(this->*chooseCenters)(branching, dsindices, indices_length, ¢ers[0], centers_length);
if (centers_length<branching) {
node->indices = dsindices;
std::sort(node->indices,node->indices+indices_length);
node->childs = NULL;
return;
}
// assign points to clusters
DistanceType cost;
computeLabels(dsindices, indices_length, ¢ers[0], centers_length, &labels[0], cost);
node->childs = pool.allocate<NodePtr>(branching);
int start = 0;
int end = start;
for (int i=0; i<branching; ++i) {
for (int j=0; j<indices_length; ++j) {
if (labels[j]==i) {
std::swap(dsindices[j],dsindices[end]);
std::swap(labels[j],labels[end]);
end++;
}
}
node->childs[i] = pool.allocate<Node>();
node->childs[i]->pivot = centers[i];
node->childs[i]->indices = NULL;
computeClustering(node->childs[i],dsindices+start, end-start, branching, level+1);
start=end;
}
}
/**
* Performs one descent in the hierarchical k-means tree. The branches not
* visited are stored in a priority queue.
*
* Params:
* node = node to explore
* result = container for the k-nearest neighbors found
* vec = query points
* checks = how many points in the dataset have been checked so far
* maxChecks = maximum dataset points to checks
*/
void findNN(NodePtr node, ResultSet<DistanceType>& result, const ElementType* vec, int& checks, int maxChecks,
Heap<BranchSt>* heap, std::vector<bool>& checked)
{
if (node->childs==NULL) {
if (checks>=maxChecks) {
if (result.full()) return;
}
for (int i=0; i<node->size; ++i) {
int index = node->indices[i];
if (!checked[index]) {
DistanceType dist = distance(dataset[index], vec, veclen_);
result.addPoint(dist, index);
checked[index] = true;
++checks;
}
}
}
else {
DistanceType* domain_distances = new DistanceType[branching_];
int best_index = 0;
domain_distances[best_index] = distance(vec, dataset[node->childs[best_index]->pivot], veclen_);
for (int i=1; i<branching_; ++i) {
domain_distances[i] = distance(vec, dataset[node->childs[i]->pivot], veclen_);
if (domain_distances[i]<domain_distances[best_index]) {
best_index = i;
}
}
for (int i=0; i<branching_; ++i) {
if (i!=best_index) {
heap->insert(BranchSt(node->childs[i],domain_distances[i]));
}
}
delete[] domain_distances;
findNN(node->childs[best_index],result,vec, checks, maxChecks, heap, checked);
}
}
private:
/**
* The dataset used by this index
*/
const Matrix<ElementType> dataset;
/**
* Parameters used by this index
*/
IndexParams params;
/**
* Number of features in the dataset.
*/
size_t size_;
/**
* Length of each feature.
*/
size_t veclen_;
/**
* The root node in the tree.
*/
NodePtr* root;
/**
* Array of indices to vectors in the dataset.
*/
int** indices;
/**
* The distance
*/
Distance distance;
/**
* Pooled memory allocator.
*
* Using a pooled memory allocator is more efficient
* than allocating memory directly when there is a large
* number small of memory allocations.
*/
PooledAllocator pool;
/**
* Memory occupied by the index.
*/
int memoryCounter;
/** index parameters */
int branching_;
int trees_;
flann_centers_init_t centers_init_;
int leaf_size_;
};
}
#endif /* OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_ */