1 /*M/////////////////////////////////////////////////////////////////////////////////////// 2 // 3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. 4 // 5 // By downloading, copying, installing or using the software you agree to this license. 6 // If you do not agree to this license, do not download, install, 7 // copy or use the software. 8 // 9 // 10 // Intel License Agreement 11 // For Open Source Computer Vision Library 12 // 13 // Copyright (C) 2008, Xavier Delacour, all rights reserved. 14 // Third party copyrights are property of their respective owners. 15 // 16 // Redistribution and use in source and binary forms, with or without modification, 17 // are permitted provided that the following conditions are met: 18 // 19 // * Redistribution's of source code must retain the above copyright notice, 20 // this list of conditions and the following disclaimer. 21 // 22 // * Redistribution's in binary form must reproduce the above copyright notice, 23 // this list of conditions and the following disclaimer in the documentation 24 // and/or other materials provided with the distribution. 25 // 26 // * The name of Intel Corporation may not be used to endorse or promote products 27 // derived from this software without specific prior written permission. 28 // 29 // This software is provided by the copyright holders and contributors "as is" and 30 // any express or implied warranties, including, but not limited to, the implied 31 // warranties of merchantability and fitness for a particular purpose are disclaimed. 32 // In no event shall the Intel Corporation or contributors be liable for any direct, 33 // indirect, incidental, special, exemplary, or consequential damages 34 // (including, but not limited to, procurement of substitute goods or services; 35 // loss of use, data, or profits; or business interruption) however caused 36 // and on any theory of liability, whether in contract, strict liability, 37 // or tort (including negligence or otherwise) arising in any way out of 38 // the use of this software, even if advised of the possibility of such damage. 39 // 40 //M*/ 41 42 // 2008-05-13, Xavier Delacour <xavier.delacour@gmail.com> 43 44 #ifndef __cv_kdtree_h__ 45 #define __cv_kdtree_h__ 46 47 #include "_cv.h" 48 49 #include <vector> 50 #include <algorithm> 51 #include <limits> 52 #include <iostream> 53 #include "assert.h" 54 #include "math.h" 55 56 // J.S. Beis and D.G. Lowe. Shape indexing using approximate nearest-neighbor search in highdimensional spaces. In Proc. IEEE Conf. Comp. Vision Patt. Recog., pages 1000--1006, 1997. http://citeseer.ist.psu.edu/beis97shape.html 57 #undef __deref 58 #undef __valuetype 59 60 template < class __valuetype, class __deref > 61 class CvKDTree { 62 public: 63 typedef __deref deref_type; 64 typedef typename __deref::scalar_type scalar_type; 65 typedef typename __deref::accum_type accum_type; 66 67 private: 68 struct node { 69 int dim; // split dimension; >=0 for nodes, -1 for leaves 70 __valuetype value; // if leaf, value of leaf 71 int left, right; // node indices of left and right branches 72 scalar_type boundary; // left if deref(value,dim)<=boundary, otherwise right 73 }; 74 typedef std::vector < node > node_array; 75 76 __deref deref; // requires operator() (__valuetype lhs,int dim) 77 78 node_array nodes; // node storage 79 int point_dim; // dimension of points (the k in kd-tree) 80 int root_node; // index of root node, -1 if empty tree 81 82 // for given set of point indices, compute dimension of highest variance 83 template < class __instype, class __valuector > dimension_of_highest_variance(__instype * first,__instype * last,__valuector ctor)84 int dimension_of_highest_variance(__instype * first, __instype * last, 85 __valuector ctor) { 86 assert(last - first > 0); 87 88 accum_type maxvar = -std::numeric_limits < accum_type >::max(); 89 int maxj = -1; 90 for (int j = 0; j < point_dim; ++j) { 91 accum_type mean = 0; 92 for (__instype * k = first; k < last; ++k) 93 mean += deref(ctor(*k), j); 94 mean /= last - first; 95 accum_type var = 0; 96 for (__instype * k = first; k < last; ++k) { 97 accum_type diff = accum_type(deref(ctor(*k), j)) - mean; 98 var += diff * diff; 99 } 100 var /= last - first; 101 102 assert(maxj != -1 || var >= maxvar); 103 104 if (var >= maxvar) { 105 maxvar = var; 106 maxj = j; 107 } 108 } 109 110 return maxj; 111 } 112 113 // given point indices and dimension, find index of median; (almost) modifies [first,last) 114 // such that points_in[first,median]<=point[median], points_in(median,last)>point[median]. 115 // implemented as partial quicksort; expected linear perf. 116 template < class __instype, class __valuector > median_partition(__instype * first,__instype * last,int dim,__valuector ctor)117 __instype * median_partition(__instype * first, __instype * last, 118 int dim, __valuector ctor) { 119 assert(last - first > 0); 120 __instype *k = first + (last - first) / 2; 121 median_partition(first, last, k, dim, ctor); 122 return k; 123 } 124 125 template < class __instype, class __valuector > 126 struct median_pr { 127 const __instype & pivot; 128 int dim; 129 __deref deref; 130 __valuector ctor; median_prCvKDTree::median_pr131 median_pr(const __instype & _pivot, int _dim, __deref _deref, __valuector _ctor) 132 : pivot(_pivot), dim(_dim), deref(_deref), ctor(_ctor) { 133 } operator ()CvKDTree::median_pr134 bool operator() (const __instype & lhs) const { 135 return deref(ctor(lhs), dim) <= deref(ctor(pivot), dim); 136 } 137 }; 138 139 template < class __instype, class __valuector > median_partition(__instype * first,__instype * last,__instype * k,int dim,__valuector ctor)140 void median_partition(__instype * first, __instype * last, 141 __instype * k, int dim, __valuector ctor) { 142 int pivot = (last - first) / 2; 143 144 std::swap(first[pivot], last[-1]); 145 __instype *middle = std::partition(first, last - 1, 146 median_pr < __instype, __valuector > 147 (last[-1], dim, deref, ctor)); 148 std::swap(*middle, last[-1]); 149 150 if (middle < k) 151 median_partition(middle + 1, last, k, dim, ctor); 152 else if (middle > k) 153 median_partition(first, middle, k, dim, ctor); 154 } 155 156 // insert given points into the tree; return created node 157 template < class __instype, class __valuector > insert(__instype * first,__instype * last,__valuector ctor)158 int insert(__instype * first, __instype * last, __valuector ctor) { 159 if (first == last) 160 return -1; 161 else { 162 163 int dim = dimension_of_highest_variance(first, last, ctor); 164 __instype *median = median_partition(first, last, dim, ctor); 165 166 __instype *split = median; 167 for (; split != last && deref(ctor(*split), dim) == 168 deref(ctor(*median), dim); ++split); 169 170 if (split == last) { // leaf 171 int nexti = -1; 172 for (--split; split >= first; --split) { 173 int i = nodes.size(); 174 node & n = *nodes.insert(nodes.end(), node()); 175 n.dim = -1; 176 n.value = ctor(*split); 177 n.left = -1; 178 n.right = nexti; 179 nexti = i; 180 } 181 182 return nexti; 183 } else { // node 184 int i = nodes.size(); 185 // note that recursive insert may invalidate this ref 186 node & n = *nodes.insert(nodes.end(), node()); 187 188 n.dim = dim; 189 n.boundary = deref(ctor(*median), dim); 190 191 int left = insert(first, split, ctor); 192 nodes[i].left = left; 193 int right = insert(split, last, ctor); 194 nodes[i].right = right; 195 196 return i; 197 } 198 } 199 } 200 201 // run to leaf; linear search for p; 202 // if found, remove paths to empty leaves on unwind remove(int * i,const __valuetype & p)203 bool remove(int *i, const __valuetype & p) { 204 if (*i == -1) 205 return false; 206 node & n = nodes[*i]; 207 bool r; 208 209 if (n.dim >= 0) { // node 210 if (deref(p, n.dim) <= n.boundary) // left 211 r = remove(&n.left, p); 212 else // right 213 r = remove(&n.right, p); 214 215 // if terminal, remove this node 216 if (n.left == -1 && n.right == -1) 217 *i = -1; 218 219 return r; 220 } else { // leaf 221 if (n.value == p) { 222 *i = n.right; 223 return true; 224 } else 225 return remove(&n.right, p); 226 } 227 } 228 229 public: 230 struct identity_ctor { operator ()CvKDTree::identity_ctor231 const __valuetype & operator() (const __valuetype & rhs) const { 232 return rhs; 233 } 234 }; 235 236 // initialize an empty tree CvKDTree(__deref _deref=__deref ())237 CvKDTree(__deref _deref = __deref()) 238 : deref(_deref), root_node(-1) { 239 } 240 // given points, initialize a balanced tree CvKDTree(__valuetype * first,__valuetype * last,int _point_dim,__deref _deref=__deref ())241 CvKDTree(__valuetype * first, __valuetype * last, int _point_dim, 242 __deref _deref = __deref()) 243 : deref(_deref) { 244 set_data(first, last, _point_dim, identity_ctor()); 245 } 246 // given points, initialize a balanced tree 247 template < class __instype, class __valuector > CvKDTree(__instype * first,__instype * last,int _point_dim,__valuector ctor,__deref _deref=__deref ())248 CvKDTree(__instype * first, __instype * last, int _point_dim, 249 __valuector ctor, __deref _deref = __deref()) 250 : deref(_deref) { 251 set_data(first, last, _point_dim, ctor); 252 } 253 set_deref(__deref _deref)254 void set_deref(__deref _deref) { 255 deref = _deref; 256 } 257 set_data(__valuetype * first,__valuetype * last,int _point_dim)258 void set_data(__valuetype * first, __valuetype * last, int _point_dim) { 259 set_data(first, last, _point_dim, identity_ctor()); 260 } 261 template < class __instype, class __valuector > set_data(__instype * first,__instype * last,int _point_dim,__valuector ctor)262 void set_data(__instype * first, __instype * last, int _point_dim, 263 __valuector ctor) { 264 point_dim = _point_dim; 265 nodes.clear(); 266 nodes.reserve(last - first); 267 root_node = insert(first, last, ctor); 268 } 269 dims() const270 int dims() const { 271 return point_dim; 272 } 273 274 // remove the given point remove(const __valuetype & p)275 bool remove(const __valuetype & p) { 276 return remove(&root_node, p); 277 } 278 print() const279 void print() const { 280 print(root_node); 281 } print(int i,int indent=0) const282 void print(int i, int indent = 0) const { 283 if (i == -1) 284 return; 285 for (int j = 0; j < indent; ++j) 286 std::cout << " "; 287 const node & n = nodes[i]; 288 if (n.dim >= 0) { 289 std::cout << "node " << i << ", left " << nodes[i].left << ", right " << 290 nodes[i].right << ", dim " << nodes[i].dim << ", boundary " << 291 nodes[i].boundary << std::endl; 292 print(n.left, indent + 3); 293 print(n.right, indent + 3); 294 } else 295 std::cout << "leaf " << i << ", value = " << nodes[i].value << std::endl; 296 } 297 298 //////////////////////////////////////////////////////////////////////////////////////// 299 // bbf search 300 public: 301 struct bbf_nn { // info on found neighbors (approx k nearest) 302 const __valuetype *p; // nearest neighbor 303 accum_type dist; // distance from d to query point bbf_nnCvKDTree::bbf_nn304 bbf_nn(const __valuetype & _p, accum_type _dist) 305 : p(&_p), dist(_dist) { 306 } operator <CvKDTree::bbf_nn307 bool operator<(const bbf_nn & rhs) const { 308 return dist < rhs.dist; 309 } 310 }; 311 typedef std::vector < bbf_nn > bbf_nn_pqueue; 312 private: 313 struct bbf_node { // info on branches not taken 314 int node; // corresponding node 315 accum_type dist; // minimum distance from bounds to query point bbf_nodeCvKDTree::bbf_node316 bbf_node(int _node, accum_type _dist) 317 : node(_node), dist(_dist) { 318 } operator <CvKDTree::bbf_node319 bool operator<(const bbf_node & rhs) const { 320 return dist > rhs.dist; 321 } 322 }; 323 typedef std::vector < bbf_node > bbf_pqueue; 324 mutable bbf_pqueue tmp_pq; 325 326 // called for branches not taken, as bbf walks to leaf; 327 // construct bbf_node given minimum distance to bounds of alternate branch pq_alternate(int alt_n,bbf_pqueue & pq,scalar_type dist) const328 void pq_alternate(int alt_n, bbf_pqueue & pq, scalar_type dist) const { 329 if (alt_n == -1) 330 return; 331 332 // add bbf_node for alternate branch in priority queue 333 pq.push_back(bbf_node(alt_n, dist)); 334 push_heap(pq.begin(), pq.end()); 335 } 336 337 // called by bbf to walk to leaf; 338 // takes one step down the tree towards query point d 339 template < class __desctype > bbf_branch(int i,const __desctype * d,bbf_pqueue & pq) const340 int bbf_branch(int i, const __desctype * d, bbf_pqueue & pq) const { 341 const node & n = nodes[i]; 342 // push bbf_node with bounds of alternate branch, then branch 343 if (d[n.dim] <= n.boundary) { // left 344 pq_alternate(n.right, pq, n.boundary - d[n.dim]); 345 return n.left; 346 } else { // right 347 pq_alternate(n.left, pq, d[n.dim] - n.boundary); 348 return n.right; 349 } 350 } 351 352 // compute euclidean distance between two points 353 template < class __desctype > distance(const __desctype * d,const __valuetype & p) const354 accum_type distance(const __desctype * d, const __valuetype & p) const { 355 accum_type dist = 0; 356 for (int j = 0; j < point_dim; ++j) { 357 accum_type diff = accum_type(d[j]) - accum_type(deref(p, j)); 358 dist += diff * diff; 359 } return (accum_type) sqrt(dist); 360 } 361 362 // called per candidate nearest neighbor; constructs new bbf_nn for 363 // candidate and adds it to priority queue of all candidates; if 364 // queue len exceeds k, drops the point furthest from query point d. 365 template < class __desctype > bbf_new_nn(bbf_nn_pqueue & nn_pq,int k,const __desctype * d,const __valuetype & p) const366 void bbf_new_nn(bbf_nn_pqueue & nn_pq, int k, 367 const __desctype * d, const __valuetype & p) const { 368 bbf_nn nn(p, distance(d, p)); 369 if ((int) nn_pq.size() < k) { 370 nn_pq.push_back(nn); 371 push_heap(nn_pq.begin(), nn_pq.end()); 372 } else if (nn_pq[0].dist > nn.dist) { 373 pop_heap(nn_pq.begin(), nn_pq.end()); 374 nn_pq.end()[-1] = nn; 375 push_heap(nn_pq.begin(), nn_pq.end()); 376 } 377 assert(nn_pq.size() < 2 || nn_pq[0].dist >= nn_pq[1].dist); 378 } 379 380 public: 381 // finds (with high probability) the k nearest neighbors of d, 382 // searching at most emax leaves/bins. 383 // ret_nn_pq is an array containing the (at most) k nearest neighbors 384 // (see bbf_nn structure def above). 385 template < class __desctype > find_nn_bbf(const __desctype * d,int k,int emax,bbf_nn_pqueue & ret_nn_pq) const386 int find_nn_bbf(const __desctype * d, 387 int k, int emax, 388 bbf_nn_pqueue & ret_nn_pq) const { 389 assert(k > 0); 390 ret_nn_pq.clear(); 391 392 if (root_node == -1) 393 return 0; 394 395 // add root_node to bbf_node priority queue; 396 // iterate while queue non-empty and emax>0 397 tmp_pq.clear(); 398 tmp_pq.push_back(bbf_node(root_node, 0)); 399 while (tmp_pq.size() && emax > 0) { 400 401 // from node nearest query point d, run to leaf 402 pop_heap(tmp_pq.begin(), tmp_pq.end()); 403 bbf_node bbf(tmp_pq.end()[-1]); 404 tmp_pq.erase(tmp_pq.end() - 1); 405 406 int i; 407 for (i = bbf.node; 408 i != -1 && nodes[i].dim >= 0; 409 i = bbf_branch(i, d, tmp_pq)); 410 411 if (i != -1) { 412 413 // add points in leaf/bin to ret_nn_pq 414 do { 415 bbf_new_nn(ret_nn_pq, k, d, nodes[i].value); 416 } while (-1 != (i = nodes[i].right)); 417 418 --emax; 419 } 420 } 421 422 tmp_pq.clear(); 423 return ret_nn_pq.size(); 424 } 425 426 //////////////////////////////////////////////////////////////////////////////////////// 427 // orthogonal range search 428 private: find_ortho_range(int i,scalar_type * bounds_min,scalar_type * bounds_max,std::vector<__valuetype> & inbounds) const429 void find_ortho_range(int i, scalar_type * bounds_min, 430 scalar_type * bounds_max, 431 std::vector < __valuetype > &inbounds) const { 432 if (i == -1) 433 return; 434 const node & n = nodes[i]; 435 if (n.dim >= 0) { // node 436 if (bounds_min[n.dim] <= n.boundary) 437 find_ortho_range(n.left, bounds_min, bounds_max, inbounds); 438 if (bounds_max[n.dim] > n.boundary) 439 find_ortho_range(n.right, bounds_min, bounds_max, inbounds); 440 } else { // leaf 441 do { 442 inbounds.push_back(nodes[i].value); 443 } while (-1 != (i = nodes[i].right)); 444 } 445 } 446 public: 447 // return all points that lie within the given bounds; inbounds is cleared find_ortho_range(scalar_type * bounds_min,scalar_type * bounds_max,std::vector<__valuetype> & inbounds) const448 int find_ortho_range(scalar_type * bounds_min, 449 scalar_type * bounds_max, 450 std::vector < __valuetype > &inbounds) const { 451 inbounds.clear(); 452 find_ortho_range(root_node, bounds_min, bounds_max, inbounds); 453 return inbounds.size(); 454 } 455 }; 456 457 #endif // __cv_kdtree_h__ 458 459 // Local Variables: 460 // mode:C++ 461 // End: 462