1 | /*********************************************************************** |
---|
2 | * Software License Agreement (BSD License) |
---|
3 | * |
---|
4 | * Copyright 2008-2009 Marius Muja (mariusm@cs.ubc.ca). All rights reserved. |
---|
5 | * Copyright 2008-2009 David G. Lowe (lowe@cs.ubc.ca). All rights reserved. |
---|
6 | * |
---|
7 | * THE BSD LICENSE |
---|
8 | * |
---|
9 | * Redistribution and use in source and binary forms, with or without |
---|
10 | * modification, are permitted provided that the following conditions |
---|
11 | * are met: |
---|
12 | * |
---|
13 | * 1. Redistributions of source code must retain the above copyright |
---|
14 | * notice, this list of conditions and the following disclaimer. |
---|
15 | * 2. Redistributions in binary form must reproduce the above copyright |
---|
16 | * notice, this list of conditions and the following disclaimer in the |
---|
17 | * documentation and/or other materials provided with the distribution. |
---|
18 | * |
---|
19 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR |
---|
20 | * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES |
---|
21 | * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. |
---|
22 | * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, |
---|
23 | * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT |
---|
24 | * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
---|
25 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
---|
26 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
---|
27 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF |
---|
28 | * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
---|
29 | *************************************************************************/ |
---|
30 | |
---|
31 | /*********************************************************************** |
---|
32 | * Author: Vincent Rabaud |
---|
33 | *************************************************************************/ |
---|
34 | |
---|
35 | #ifndef FLANN_LSH_INDEX_H_ |
---|
36 | #define FLANN_LSH_INDEX_H_ |
---|
37 | |
---|
38 | #include <algorithm> |
---|
39 | #include <cassert> |
---|
40 | #include <cstring> |
---|
41 | #include <map> |
---|
42 | #include <vector> |
---|
43 | |
---|
44 | #include "flann/general.h" |
---|
45 | #include "flann/algorithms/nn_index.h" |
---|
46 | #include "flann/util/matrix.h" |
---|
47 | #include "flann/util/result_set.h" |
---|
48 | #include "flann/util/heap.h" |
---|
49 | #include "flann/util/lsh_table.h" |
---|
50 | #include "flann/util/allocator.h" |
---|
51 | #include "flann/util/random.h" |
---|
52 | #include "flann/util/saving.h" |
---|
53 | |
---|
54 | namespace flann |
---|
55 | { |
---|
56 | |
---|
57 | struct LshIndexParams : public IndexParams |
---|
58 | { |
---|
59 | LshIndexParams(unsigned int table_number = 12, unsigned int key_size = 20, unsigned int multi_probe_level = 2) |
---|
60 | { |
---|
61 | (* this)["algorithm"] = FLANN_INDEX_LSH; |
---|
62 | // The number of hash tables to use |
---|
63 | (*this)["table_number"] = table_number; |
---|
64 | // The length of the key in the hash tables |
---|
65 | (*this)["key_size"] = key_size; |
---|
66 | // Number of levels to use in multi-probe (0 for standard LSH) |
---|
67 | (*this)["multi_probe_level"] = multi_probe_level; |
---|
68 | } |
---|
69 | }; |
---|
70 | |
---|
71 | /** |
---|
72 | * Randomized kd-tree index |
---|
73 | * |
---|
74 | * Contains the k-d trees and other information for indexing a set of points |
---|
75 | * for nearest-neighbor matching. |
---|
76 | */ |
---|
77 | template<typename Distance> |
---|
78 | class LshIndex : public NNIndex<Distance> |
---|
79 | { |
---|
80 | public: |
---|
81 | typedef typename Distance::ElementType ElementType; |
---|
82 | typedef typename Distance::ResultType DistanceType; |
---|
83 | |
---|
84 | /** Constructor |
---|
85 | * @param input_data dataset with the input features |
---|
86 | * @param params parameters passed to the LSH algorithm |
---|
87 | * @param d the distance used |
---|
88 | */ |
---|
89 | LshIndex(const Matrix<ElementType>& input_data, const IndexParams& params = LshIndexParams(), |
---|
90 | Distance d = Distance()) : |
---|
91 | dataset_(input_data), index_params_(params), distance_(d) |
---|
92 | { |
---|
93 | table_number_ = get_param<unsigned int>(index_params_,"table_number",12); |
---|
94 | key_size_ = get_param<unsigned int>(index_params_,"key_size",20); |
---|
95 | multi_probe_level_ = get_param<unsigned int>(index_params_,"multi_probe_level",2); |
---|
96 | |
---|
97 | feature_size_ = dataset_.cols; |
---|
98 | fill_xor_mask(0, key_size_, multi_probe_level_, xor_masks_); |
---|
99 | } |
---|
100 | |
---|
101 | |
---|
102 | LshIndex(const LshIndex&); |
---|
103 | LshIndex& operator=(const LshIndex&); |
---|
104 | |
---|
105 | /** |
---|
106 | * Builds the index |
---|
107 | */ |
---|
108 | void buildIndex() |
---|
109 | { |
---|
110 | tables_.resize(table_number_); |
---|
111 | for (unsigned int i = 0; i < table_number_; ++i) { |
---|
112 | lsh::LshTable<ElementType>& table = tables_[i]; |
---|
113 | table = lsh::LshTable<ElementType>(feature_size_, key_size_); |
---|
114 | |
---|
115 | // Add the features to the table |
---|
116 | table.add(dataset_); |
---|
117 | } |
---|
118 | } |
---|
119 | |
---|
120 | flann_algorithm_t getType() const |
---|
121 | { |
---|
122 | return FLANN_INDEX_LSH; |
---|
123 | } |
---|
124 | |
---|
125 | |
---|
126 | void saveIndex(FILE* stream) |
---|
127 | { |
---|
128 | save_value(stream,table_number_); |
---|
129 | save_value(stream,key_size_); |
---|
130 | save_value(stream,multi_probe_level_); |
---|
131 | save_value(stream, dataset_); |
---|
132 | } |
---|
133 | |
---|
134 | void loadIndex(FILE* stream) |
---|
135 | { |
---|
136 | load_value(stream, table_number_); |
---|
137 | load_value(stream, key_size_); |
---|
138 | load_value(stream, multi_probe_level_); |
---|
139 | load_value(stream, dataset_); |
---|
140 | // Building the index is so fast we can afford not storing it |
---|
141 | buildIndex(); |
---|
142 | |
---|
143 | index_params_["algorithm"] = getType(); |
---|
144 | index_params_["table_number"] = table_number_; |
---|
145 | index_params_["key_size"] = key_size_; |
---|
146 | index_params_["multi_probe_level"] = multi_probe_level_; |
---|
147 | } |
---|
148 | |
---|
149 | /** |
---|
150 | * Returns size of index. |
---|
151 | */ |
---|
152 | size_t size() const |
---|
153 | { |
---|
154 | return dataset_.rows; |
---|
155 | } |
---|
156 | |
---|
157 | /** |
---|
158 | * Returns the length of an index feature. |
---|
159 | */ |
---|
160 | size_t veclen() const |
---|
161 | { |
---|
162 | return feature_size_; |
---|
163 | } |
---|
164 | |
---|
165 | /** |
---|
166 | * Computes the index memory usage |
---|
167 | * Returns: memory used by the index |
---|
168 | */ |
---|
169 | int usedMemory() const |
---|
170 | { |
---|
171 | return dataset_.rows * sizeof(int); |
---|
172 | } |
---|
173 | |
---|
174 | |
---|
175 | IndexParams getParameters() const |
---|
176 | { |
---|
177 | return index_params_; |
---|
178 | } |
---|
179 | |
---|
180 | /** |
---|
181 | * \brief Perform k-nearest neighbor search |
---|
182 | * \param[in] queries The query points for which to find the nearest neighbors |
---|
183 | * \param[out] indices The indices of the nearest neighbors found |
---|
184 | * \param[out] dists Distances to the nearest neighbors found |
---|
185 | * \param[in] knn Number of nearest neighbors to return |
---|
186 | * \param[in] params Search parameters |
---|
187 | */ |
---|
188 | virtual int knnSearch(const Matrix<ElementType>& queries, |
---|
189 | Matrix<int>& indices, |
---|
190 | Matrix<DistanceType>& dists, |
---|
191 | size_t knn, |
---|
192 | const SearchParams& params) |
---|
193 | { |
---|
194 | assert(queries.cols == veclen()); |
---|
195 | assert(indices.rows >= queries.rows); |
---|
196 | assert(dists.rows >= queries.rows); |
---|
197 | assert(indices.cols >= knn); |
---|
198 | assert(dists.cols >= knn); |
---|
199 | |
---|
200 | int count = 0; |
---|
201 | if (params.use_heap==FLANN_True) { |
---|
202 | KNNUniqueResultSet<DistanceType> resultSet(knn); |
---|
203 | for (size_t i = 0; i < queries.rows; i++) { |
---|
204 | resultSet.clear(); |
---|
205 | findNeighbors(resultSet, queries[i], params); |
---|
206 | resultSet.copy(indices[i], dists[i], knn, params.sorted); |
---|
207 | count += resultSet.size(); |
---|
208 | } |
---|
209 | } |
---|
210 | else { |
---|
211 | KNNResultSet<DistanceType> resultSet(knn); |
---|
212 | for (size_t i = 0; i < queries.rows; i++) { |
---|
213 | resultSet.clear(); |
---|
214 | findNeighbors(resultSet, queries[i], params); |
---|
215 | resultSet.copy(indices[i], dists[i], knn, params.sorted); |
---|
216 | count += resultSet.size(); |
---|
217 | } |
---|
218 | } |
---|
219 | |
---|
220 | return count; |
---|
221 | } |
---|
222 | |
---|
223 | /** |
---|
224 | * \brief Perform k-nearest neighbor search |
---|
225 | * \param[in] queries The query points for which to find the nearest neighbors |
---|
226 | * \param[out] indices The indices of the nearest neighbors found |
---|
227 | * \param[out] dists Distances to the nearest neighbors found |
---|
228 | * \param[in] knn Number of nearest neighbors to return |
---|
229 | * \param[in] params Search parameters |
---|
230 | */ |
---|
231 | virtual int knnSearch(const Matrix<ElementType>& queries, |
---|
232 | std::vector< std::vector<int> >& indices, |
---|
233 | std::vector<std::vector<DistanceType> >& dists, |
---|
234 | size_t knn, |
---|
235 | const SearchParams& params) |
---|
236 | { |
---|
237 | assert(queries.cols == veclen()); |
---|
238 | if (indices.size() < queries.rows ) indices.resize(queries.rows); |
---|
239 | if (dists.size() < queries.rows ) dists.resize(queries.rows); |
---|
240 | |
---|
241 | int count = 0; |
---|
242 | if (params.use_heap==FLANN_True) { |
---|
243 | KNNUniqueResultSet<DistanceType> resultSet(knn); |
---|
244 | for (size_t i = 0; i < queries.rows; i++) { |
---|
245 | resultSet.clear(); |
---|
246 | findNeighbors(resultSet, queries[i], params); |
---|
247 | size_t n = std::min(resultSet.size(), knn); |
---|
248 | indices[i].resize(n); |
---|
249 | dists[i].resize(n); |
---|
250 | resultSet.copy(&indices[i][0], &dists[i][0], n, params.sorted); |
---|
251 | count += n; |
---|
252 | } |
---|
253 | } |
---|
254 | else { |
---|
255 | KNNResultSet<DistanceType> resultSet(knn); |
---|
256 | for (size_t i = 0; i < queries.rows; i++) { |
---|
257 | resultSet.clear(); |
---|
258 | findNeighbors(resultSet, queries[i], params); |
---|
259 | size_t n = std::min(resultSet.size(), knn); |
---|
260 | indices[i].resize(n); |
---|
261 | dists[i].resize(n); |
---|
262 | resultSet.copy(&indices[i][0], &dists[i][0], n, params.sorted); |
---|
263 | count += n; |
---|
264 | } |
---|
265 | } |
---|
266 | |
---|
267 | return count; |
---|
268 | } |
---|
269 | |
---|
270 | /** |
---|
271 | * Find set of nearest neighbors to vec. Their indices are stored inside |
---|
272 | * the result object. |
---|
273 | * |
---|
274 | * Params: |
---|
275 | * result = the result object in which the indices of the nearest-neighbors are stored |
---|
276 | * vec = the vector for which to search the nearest neighbors |
---|
277 | * maxCheck = the maximum number of restarts (in a best-bin-first manner) |
---|
278 | */ |
---|
279 | void findNeighbors(ResultSet<DistanceType>& result, const ElementType* vec, const SearchParams& /*searchParams*/) |
---|
280 | { |
---|
281 | getNeighbors(vec, result); |
---|
282 | } |
---|
283 | |
---|
284 | private: |
---|
285 | /** Defines the comparator on score and index |
---|
286 | */ |
---|
287 | typedef std::pair<float, unsigned int> ScoreIndexPair; |
---|
288 | struct SortScoreIndexPairOnSecond |
---|
289 | { |
---|
290 | bool operator()(const ScoreIndexPair& left, const ScoreIndexPair& right) const |
---|
291 | { |
---|
292 | return left.second < right.second; |
---|
293 | } |
---|
294 | }; |
---|
295 | |
---|
296 | /** Fills the different xor masks to use when getting the neighbors in multi-probe LSH |
---|
297 | * @param key the key we build neighbors from |
---|
298 | * @param lowest_index the lowest index of the bit set |
---|
299 | * @param level the multi-probe level we are at |
---|
300 | * @param xor_masks all the xor mask |
---|
301 | */ |
---|
302 | void fill_xor_mask(lsh::BucketKey key, int lowest_index, unsigned int level, |
---|
303 | std::vector<lsh::BucketKey>& xor_masks) |
---|
304 | { |
---|
305 | xor_masks.push_back(key); |
---|
306 | if (level == 0) return; |
---|
307 | for (int index = lowest_index - 1; index >= 0; --index) { |
---|
308 | // Create a new key |
---|
309 | lsh::BucketKey new_key = key | (1 << index); |
---|
310 | fill_xor_mask(new_key, index, level - 1, xor_masks); |
---|
311 | } |
---|
312 | } |
---|
313 | |
---|
314 | /** Performs the approximate nearest-neighbor search. |
---|
315 | * @param vec the feature to analyze |
---|
316 | * @param do_radius flag indicating if we check the radius too |
---|
317 | * @param radius the radius if it is a radius search |
---|
318 | * @param do_k flag indicating if we limit the number of nn |
---|
319 | * @param k_nn the number of nearest neighbors |
---|
320 | * @param checked_average used for debugging |
---|
321 | */ |
---|
322 | void getNeighbors(const ElementType* vec, bool do_radius, float radius, bool do_k, unsigned int k_nn, |
---|
323 | float& checked_average) |
---|
324 | { |
---|
325 | static std::vector<ScoreIndexPair> score_index_heap; |
---|
326 | |
---|
327 | if (do_k) { |
---|
328 | unsigned int worst_score = std::numeric_limits<unsigned int>::max(); |
---|
329 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin(); |
---|
330 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end(); |
---|
331 | for (; table != table_end; ++table) { |
---|
332 | size_t key = table->getKey(vec); |
---|
333 | std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin(); |
---|
334 | std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end(); |
---|
335 | for (; xor_mask != xor_mask_end; ++xor_mask) { |
---|
336 | size_t sub_key = key ^ (*xor_mask); |
---|
337 | const lsh::Bucket* bucket = table->getBucketFromKey(sub_key); |
---|
338 | if (bucket == 0) continue; |
---|
339 | |
---|
340 | // Go over each descriptor index |
---|
341 | std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin(); |
---|
342 | std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end(); |
---|
343 | DistanceType hamming_distance; |
---|
344 | |
---|
345 | // Process the rest of the candidates |
---|
346 | for (; training_index < last_training_index; ++training_index) { |
---|
347 | hamming_distance = distance_(vec, dataset_[*training_index], dataset_.cols); |
---|
348 | |
---|
349 | if (hamming_distance < worst_score) { |
---|
350 | // Insert the new element |
---|
351 | score_index_heap.push_back(ScoreIndexPair(hamming_distance, training_index)); |
---|
352 | std::push_heap(score_index_heap.begin(), score_index_heap.end()); |
---|
353 | |
---|
354 | if (score_index_heap.size() > (unsigned int)k_nn) { |
---|
355 | // Remove the highest distance value as we have too many elements |
---|
356 | std::pop_heap(score_index_heap.begin(), score_index_heap.end()); |
---|
357 | score_index_heap.pop_back(); |
---|
358 | // Keep track of the worst score |
---|
359 | worst_score = score_index_heap.front().first; |
---|
360 | } |
---|
361 | } |
---|
362 | } |
---|
363 | } |
---|
364 | } |
---|
365 | } |
---|
366 | else { |
---|
367 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin(); |
---|
368 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end(); |
---|
369 | for (; table != table_end; ++table) { |
---|
370 | size_t key = table->getKey(vec); |
---|
371 | std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin(); |
---|
372 | std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end(); |
---|
373 | for (; xor_mask != xor_mask_end; ++xor_mask) { |
---|
374 | size_t sub_key = key ^ (*xor_mask); |
---|
375 | const lsh::Bucket* bucket = table->getBucketFromKey(sub_key); |
---|
376 | if (bucket == 0) continue; |
---|
377 | |
---|
378 | // Go over each descriptor index |
---|
379 | std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin(); |
---|
380 | std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end(); |
---|
381 | DistanceType hamming_distance; |
---|
382 | |
---|
383 | // Process the rest of the candidates |
---|
384 | for (; training_index < last_training_index; ++training_index) { |
---|
385 | // Compute the Hamming distance |
---|
386 | hamming_distance = distance_(vec, dataset_[*training_index], dataset_.cols); |
---|
387 | if (hamming_distance < radius) score_index_heap.push_back(ScoreIndexPair(hamming_distance, training_index)); |
---|
388 | } |
---|
389 | } |
---|
390 | } |
---|
391 | } |
---|
392 | } |
---|
393 | |
---|
394 | /** Performs the approximate nearest-neighbor search. |
---|
395 | * This is a slower version than the above as it uses the ResultSet |
---|
396 | * @param vec the feature to analyze |
---|
397 | */ |
---|
398 | void getNeighbors(const ElementType* vec, ResultSet<DistanceType>& result) |
---|
399 | { |
---|
400 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table = tables_.begin(); |
---|
401 | typename std::vector<lsh::LshTable<ElementType> >::const_iterator table_end = tables_.end(); |
---|
402 | for (; table != table_end; ++table) { |
---|
403 | size_t key = table->getKey(vec); |
---|
404 | std::vector<lsh::BucketKey>::const_iterator xor_mask = xor_masks_.begin(); |
---|
405 | std::vector<lsh::BucketKey>::const_iterator xor_mask_end = xor_masks_.end(); |
---|
406 | for (; xor_mask != xor_mask_end; ++xor_mask) { |
---|
407 | size_t sub_key = key ^ (*xor_mask); |
---|
408 | const lsh::Bucket* bucket = table->getBucketFromKey(sub_key); |
---|
409 | if (bucket == 0) continue; |
---|
410 | |
---|
411 | // Go over each descriptor index |
---|
412 | std::vector<lsh::FeatureIndex>::const_iterator training_index = bucket->begin(); |
---|
413 | std::vector<lsh::FeatureIndex>::const_iterator last_training_index = bucket->end(); |
---|
414 | DistanceType hamming_distance; |
---|
415 | |
---|
416 | // Process the rest of the candidates |
---|
417 | for (; training_index < last_training_index; ++training_index) { |
---|
418 | // Compute the Hamming distance |
---|
419 | hamming_distance = distance_(vec, dataset_[*training_index], dataset_.cols); |
---|
420 | result.addPoint(hamming_distance, *training_index); |
---|
421 | } |
---|
422 | } |
---|
423 | } |
---|
424 | } |
---|
425 | |
---|
426 | /** The different hash tables */ |
---|
427 | std::vector<lsh::LshTable<ElementType> > tables_; |
---|
428 | |
---|
429 | /** The data the LSH tables where built from */ |
---|
430 | Matrix<ElementType> dataset_; |
---|
431 | |
---|
432 | /** The size of the features (as ElementType[]) */ |
---|
433 | unsigned int feature_size_; |
---|
434 | |
---|
435 | IndexParams index_params_; |
---|
436 | |
---|
437 | /** table number */ |
---|
438 | unsigned int table_number_; |
---|
439 | /** key size */ |
---|
440 | unsigned int key_size_; |
---|
441 | /** How far should we look for neighbors in multi-probe LSH */ |
---|
442 | unsigned int multi_probe_level_; |
---|
443 | |
---|
444 | /** The XOR masks to apply to a key to get the neighboring buckets */ |
---|
445 | std::vector<lsh::BucketKey> xor_masks_; |
---|
446 | |
---|
447 | Distance distance_; |
---|
448 | }; |
---|
449 | } |
---|
450 | |
---|
451 | #endif //FLANN_LSH_INDEX_H_ |
---|