• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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