lsh_index.h

00001 /***********************************************************************
00002  * Software License Agreement (BSD License)
00003  *
00004  * Copyright 2008-2009  Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
00005  * Copyright 2008-2009  David G. Lowe (lowe@cs.ubc.ca). All rights reserved.
00006  *
00007  * THE BSD LICENSE
00008  *
00009  * Redistribution and use in source and binary forms, with or without
00010  * modification, are permitted provided that the following conditions
00011  * are met:
00012  *
00013  * 1. Redistributions of source code must retain the above copyright
00014  *    notice, this list of conditions and the following disclaimer.
00015  * 2. Redistributions in binary form must reproduce the above copyright
00016  *    notice, this list of conditions and the following disclaimer in the
00017  *    documentation and/or other materials provided with the distribution.
00018  *
00019  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
00020  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
00021  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
00022  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
00023  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
00024  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
00025  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
00026  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
00027  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
00028  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
00029  *************************************************************************/
00030 
00031 /***********************************************************************
00032  * Author: Vincent Rabaud
00033  *************************************************************************/
00034 
00035 #ifndef OPENCV_FLANN_LSH_INDEX_H_
00036 #define OPENCV_FLANN_LSH_INDEX_H_
00037 
00038 #include <algorithm>
00039 #include <cassert>
00040 #include <cstring>
00041 #include <map>
00042 #include <vector>
00043 
00044 #include "general.h"
00045 #include "nn_index.h"
00046 #include "matrix.h"
00047 #include "result_set.h"
00048 #include "heap.h"
00049 #include "lsh_table.h"
00050 #include "allocator.h"
00051 #include "random.h"
00052 #include "saving.h"
00053 
00054 namespace cvflann
00055 {
00056 
00057 struct LshIndexParams : public IndexParams
00058 {
00059     LshIndexParams(unsigned int table_number = 12, unsigned int key_size = 20, unsigned int multi_probe_level = 2)
00060     {
00061         (* this)["algorithm"] = FLANN_INDEX_LSH;
00062         // The number of hash tables to use
00063         (*this)["table_number"] = table_number;
00064         // The length of the key in the hash tables
00065         (*this)["key_size"] = key_size;
00066         // Number of levels to use in multi-probe (0 for standard LSH)
00067         (*this)["multi_probe_level"] = multi_probe_level;
00068     }
00069 };
00070 
00071 /**
00072  * Randomized kd-tree index
00073  *
00074  * Contains the k-d trees and other information for indexing a set of points
00075  * for nearest-neighbor matching.
00076  */
00077 template<typename Distance>
00078 class LshIndex : public NNIndex<Distance>
00079 {
00080 public:
00081     typedef typename Distance::ElementType ElementType;
00082     typedef typename Distance::ResultType DistanceType;
00083 
00084     /** Constructor
00085      * @param input_data dataset with the input features
00086      * @param params parameters passed to the LSH algorithm
00087      * @param d the distance used
00088      */
00089     LshIndex(const Matrix<ElementType> & input_data, const IndexParams& params = LshIndexParams(),
00090              Distance d = Distance()) :
00091         dataset_(input_data), index_params_(params), distance_(d)
00092     {
00093         // cv::flann::IndexParams sets integer params as 'int', so it is used with get_param
00094         // in place of 'unsigned int'
00095         table_number_ = (unsigned int)get_param<int>(index_params_,"table_number",12);
00096         key_size_ = (unsigned int)get_param<int>(index_params_,"key_size",20);
00097         multi_probe_level_ = (unsigned int)get_param<int>(index_params_,"multi_probe_level",2);
00098 
00099         feature_size_ = (unsigned)dataset_.cols;
00100         fill_xor_mask(0, key_size_, multi_probe_level_, xor_masks_);
00101     }
00102 
00103 
00104     LshIndex(const LshIndex&);
00105     LshIndex& operator=(const LshIndex&);
00106 
00107     /**
00108      * Builds the index
00109      */
00110     void buildIndex()
00111     {
00112         tables_.resize(table_number_);
00113         for (unsigned int i = 0; i < table_number_; ++i) {
00114             lsh::LshTable<ElementType>& table = tables_[i];
00115             table = lsh::LshTable<ElementType>(feature_size_, key_size_);
00116 
00117             // Add the features to the table
00118             table.add(dataset_);
00119         }
00120     }
00121 
00122     flann_algorithm_t getType () const
00123     {
00124         return FLANN_INDEX_LSH;
00125     }
00126 
00127 
00128     void saveIndex(FILE* stream)
00129     {
00130         save_value(stream,table_number_);
00131         save_value(stream,key_size_);
00132         save_value(stream,multi_probe_level_);
00133         save_value(stream, dataset_);
00134     }
00135 
00136     void loadIndex(FILE* stream)
00137     {
00138         load_value(stream, table_number_);
00139         load_value(stream, key_size_);
00140         load_value(stream, multi_probe_level_);
00141         load_value(stream, dataset_);
00142         // Building the index is so fast we can afford not storing it
00143         buildIndex();
00144 
00145         index_params_["algorithm"] = getType ();
00146         index_params_["table_number"] = table_number_;
00147         index_params_["key_size"] = key_size_;
00148         index_params_["multi_probe_level"] = multi_probe_level_;
00149     }
00150 
00151     /**
00152      *  Returns size of index.
00153      */
00154     size_t size() const
00155     {
00156         return dataset_.rows;
00157     }
00158 
00159     /**
00160      * Returns the length of an index feature.
00161      */
00162     size_t veclen() const
00163     {
00164         return feature_size_;
00165     }
00166 
00167     /**
00168      * Computes the index memory usage
00169      * Returns: memory used by the index
00170      */
00171     int usedMemory() const
00172     {
00173         return (int)(dataset_.rows * sizeof(int));
00174     }
00175 
00176 
00177     IndexParams getParameters () const
00178     {
00179         return index_params_;
00180     }
00181 
00182     /**
00183      * \brief Perform k-nearest neighbor search
00184      * \param[in] queries The query points for which to find the nearest neighbors
00185      * \param[out] indices The indices of the nearest neighbors found
00186      * \param[out] dists Distances to the nearest neighbors found
00187      * \param[in] knn Number of nearest neighbors to return
00188      * \param[in] params Search parameters
00189      */
00190     virtual void knnSearch(const Matrix<ElementType> & queries, Matrix<int> & indices, Matrix<DistanceType>& dists, int knn, const SearchParams& params)
00191     {
00192         assert(queries.cols == veclen());
00193         assert(indices.rows >= queries.rows);
00194         assert(dists.rows >= queries.rows);
00195         assert(int(indices.cols) >= knn);
00196         assert(int(dists.cols) >= knn);
00197 
00198 
00199         KNNUniqueResultSet<DistanceType> resultSet(knn);
00200         for (size_t i = 0; i < queries.rows; i++) {
00201             resultSet.clear();
00202             std::fill_n(indices[i], knn, -1);
00203             std::fill_n(dists[i], knn, std::numeric_limits<DistanceType>::max());
00204             findNeighbors(resultSet, queries[i], params);
00205             if (get_param(params,"sorted",true)) resultSet.sortAndCopy(indices[i], dists[i], knn);
00206             else resultSet.copy(indices[i], dists[i], knn);
00207         }
00208     }
00209 
00210 
00211     /**
00212      * Find set of nearest neighbors to vec. Their indices are stored inside
00213      * the result object.
00214      *
00215      * Params:
00216      *     result = the result object in which the indices of the nearest-neighbors are stored
00217      *     vec = the vector for which to search the nearest neighbors
00218      *     maxCheck = the maximum number of restarts (in a best-bin-first manner)
00219      */
00220     void findNeighbors(ResultSet<DistanceType>& result, const ElementType* vec, const SearchParams& /*searchParams*/)
00221     {
00222         getNeighbors(vec, result);
00223     }
00224 
00225 private:
00226     /** Defines the comparator on score and index
00227      */
00228     typedef std::pair<float, unsigned int> ScoreIndexPair;
00229     struct SortScoreIndexPairOnSecond
00230     {
00231         bool operator()(const ScoreIndexPair& left, const ScoreIndexPair& right) const
00232         {
00233             return left.second < right.second;
00234         }
00235     };
00236 
00237     /** Fills the different xor masks to use when getting the neighbors in multi-probe LSH
00238      * @param key the key we build neighbors from
00239      * @param lowest_index the lowest index of the bit set
00240      * @param level the multi-probe level we are at
00241      * @param xor_masks all the xor mask
00242      */
00243     void fill_xor_mask(lsh::BucketKey key, int lowest_index, unsigned int level,
00244                        std::vector<lsh::BucketKey>& xor_masks)
00245     {
00246         xor_masks.push_back(key);
00247         if (level == 0) return;
00248         for (int index = lowest_index - 1; index >= 0; --index) {
00249             // Create a new key
00250             lsh::BucketKey new_key = key | (1 << index);
00251             fill_xor_mask(new_key, index, level - 1, xor_masks);
00252         }
00253     }
00254 
00255     /** Performs the approximate nearest-neighbor search.
00256      * @param vec the feature to analyze
00257      * @param do_radius flag indicating if we check the radius too
00258      * @param radius the radius if it is a radius search
00259      * @param do_k flag indicating if we limit the number of nn
00260      * @param k_nn the number of nearest neighbors
00261      * @param checked_average used for debugging
00262      */
00263     void getNeighbors(const ElementType* vec, bool /*do_radius*/, float radius, bool do_k, unsigned int k_nn,
00264                       float& /*checked_average*/)
00265     {
00266         static std::vector<ScoreIndexPair> score_index_heap;
00267 
00268         if (do_k) {
00269             unsigned int worst_score = std::numeric_limits<unsigned int>::max();
00270             typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin();
00271             typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end();
00272             for (; table != table_end; ++table) {
00273                 size_t key = table->getKey(vec);
00274                 std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin();
00275                 std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end();
00276                 for (; xor_mask != xor_mask_end; ++xor_mask) {
00277                     size_t sub_key = key ^ (*xor_mask);
00278                     const lsh::Bucket* bucket = table->getBucketFromKey(sub_key);
00279                     if (bucket == 0) continue;
00280 
00281                     // Go over each descriptor index
00282                     std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin();
00283                     std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end();
00284                     DistanceType hamming_distance;
00285 
00286                     // Process the rest of the candidates
00287                     for (; training_index < last_training_index; ++training_index) {
00288                         hamming_distance = distance_(vec, dataset_[*training_index], dataset_.cols);
00289 
00290                         if (hamming_distance < worst_score) {
00291                             // Insert the new element
00292                             score_index_heap.push_back(ScoreIndexPair(hamming_distance, training_index));
00293                             std::push_heap(score_index_heap.begin(), score_index_heap.end());
00294 
00295                             if (score_index_heap.size() > (unsigned int)k_nn) {
00296                                 // Remove the highest distance value as we have too many elements
00297                                 std::pop_heap(score_index_heap.begin(), score_index_heap.end());
00298                                 score_index_heap.pop_back();
00299                                 // Keep track of the worst score
00300                                 worst_score = score_index_heap.front().first;
00301                             }
00302                         }
00303                     }
00304                 }
00305             }
00306         }
00307         else {
00308             typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin();
00309             typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end();
00310             for (; table != table_end; ++table) {
00311                 size_t key = table->getKey(vec);
00312                 std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin();
00313                 std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end();
00314                 for (; xor_mask != xor_mask_end; ++xor_mask) {
00315                     size_t sub_key = key ^ (*xor_mask);
00316                     const lsh::Bucket* bucket = table->getBucketFromKey(sub_key);
00317                     if (bucket == 0) continue;
00318 
00319                     // Go over each descriptor index
00320                     std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin();
00321                     std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end();
00322                     DistanceType hamming_distance;
00323 
00324                     // Process the rest of the candidates
00325                     for (; training_index < last_training_index; ++training_index) {
00326                         // Compute the Hamming distance
00327                         hamming_distance = distance_(vec, dataset_[*training_index], dataset_.cols);
00328                         if (hamming_distance < radius) score_index_heap.push_back(ScoreIndexPair(hamming_distance, training_index));
00329                     }
00330                 }
00331             }
00332         }
00333     }
00334 
00335     /** Performs the approximate nearest-neighbor search.
00336      * This is a slower version than the above as it uses the ResultSet
00337      * @param vec the feature to analyze
00338      */
00339     void getNeighbors(const ElementType* vec, ResultSet<DistanceType>& result)
00340     {
00341         typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin();
00342         typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end();
00343         for (; table != table_end; ++table) {
00344             size_t key = table->getKey(vec);
00345             std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin();
00346             std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end();
00347             for (; xor_mask != xor_mask_end; ++xor_mask) {
00348                 size_t sub_key = key ^ (*xor_mask);
00349                 const lsh::Bucket* bucket = table->getBucketFromKey((lsh::BucketKey)sub_key);
00350                 if (bucket == 0) continue;
00351 
00352                 // Go over each descriptor index
00353                 std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin();
00354                 std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end();
00355                 DistanceType hamming_distance;
00356 
00357                 // Process the rest of the candidates
00358                 for (; training_index < last_training_index; ++training_index) {
00359                     // Compute the Hamming distance
00360                     hamming_distance = distance_(vec, dataset_[*training_index], (int)dataset_.cols);
00361                     result.addPoint(hamming_distance, *training_index);
00362                 }
00363             }
00364         }
00365     }
00366 
00367     /** The different hash tables */
00368     std::vector<lsh::LshTable<ElementType> > tables_;
00369 
00370     /** The data the LSH tables where built from */
00371     Matrix<ElementType> dataset_;
00372 
00373     /** The size of the features (as ElementType[]) */
00374     unsigned int feature_size_;
00375 
00376     IndexParams index_params_;
00377 
00378     /** table number */
00379     unsigned int table_number_;
00380     /** key size */
00381     unsigned int key_size_;
00382     /** How far should we look for neighbors in multi-probe LSH */
00383     unsigned int multi_probe_level_;
00384 
00385     /** The XOR masks to apply to a key to get the neighboring buckets */
00386     std::vector<lsh::BucketKey> xor_masks_;
00387 
00388     Distance distance_;
00389 };
00390 }
00391 
00392 #endif //OPENCV_FLANN_LSH_INDEX_H_