1 /* Original code has been submitted by Liu Liu. Here is the copyright.
2 ----------------------------------------------------------------------------------
3  * An OpenCV Implementation of SURF
4  * Further Information Refer to "SURF: Speed-Up Robust Feature"
5  * Author: Liu Liu
6  * liuliu.1987+opencv@gmail.com
7  *
8  * There are still serveral lacks for this experimental implementation:
9  * 1.The interpolation of sub-pixel mentioned in article was not implemented yet;
10  * 2.A comparision with original libSurf.so shows that the hessian detector is not a 100% match to their implementation;
11  * 3.Due to above reasons, I recommanded the original one for study and reuse;
12  *
13  * However, the speed of this implementation is something comparable to original one.
14  *
15  * Copyright© 2008, Liu Liu All rights reserved.
16  *
17  * Redistribution and use in source and binary forms, with or
18  * without modification, are permitted provided that the following
19  * conditions are met:
20  * 	Redistributions of source code must retain the above
21  * 	copyright notice, this list of conditions and the following
22  * 	disclaimer.
23  * 	Redistributions in binary form must reproduce the above
24  * 	copyright notice, this list of conditions and the following
25  * 	disclaimer in the documentation and/or other materials
26  * 	provided with the distribution.
27  * 	The name of Contributor may not be used to endorse or
28  * 	promote products derived from this software without
29  * 	specific prior written permission.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
32  * CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
33  * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
34  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
35  * DISCLAIMED. IN NO EVENT SHALL THE CONTRIBUTORS BE LIABLE FOR ANY
36  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
37  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
38  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
39  * OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
40  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
41  * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
42  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY
43  * OF SUCH DAMAGE.
44  */
45 
46 /*
47    The following changes have been made, comparing to the original contribution:
48    1. A lot of small optimizations, less memory allocations, got rid of global buffers
49    2. Reversed order of cvGetQuadrangleSubPix and cvResize calls; probably less accurate, but much faster
50    3. The descriptor computing part (which is most expensive) is threaded using OpenMP
51    (subpixel-accurate keypoint localization and scale estimation are still TBD)
52 */
53 
54 #include "_cv.h"
55 
cvSURFParams(double threshold,int extended)56 CvSURFParams cvSURFParams(double threshold, int extended)
57 {
58     CvSURFParams params;
59     params.hessianThreshold = threshold;
60     params.extended = extended;
61     params.nOctaves = 3;
62     params.nOctaveLayers = 4;
63     return params;
64 }
65 
66 struct CvSurfHF
67 {
68     int p0, p1, p2, p3;
69     float w;
70 };
71 
72 CV_INLINE float
icvCalcHaarPattern(const int * origin,const CvSurfHF * f,int n)73 icvCalcHaarPattern( const int* origin, const CvSurfHF* f, int n )
74 {
75     double d = 0;
76     for( int k = 0; k < n; k++ )
77         d += (origin[f[k].p0] + origin[f[k].p3] - origin[f[k].p1] - origin[f[k].p2])*f[k].w;
78     return (float)d;
79 }
80 
81 static void
icvResizeHaarPattern(const int src[][5],CvSurfHF * dst,int n,int oldSize,int newSize,int widthStep)82 icvResizeHaarPattern( const int src[][5], CvSurfHF* dst, int n, int oldSize, int newSize, int widthStep )
83 {
84     for( int k = 0; k < n; k++ )
85     {
86         int dx1 = src[k][0]*newSize/oldSize;
87         int dy1 = src[k][1]*newSize/oldSize;
88         int dx2 = src[k][2]*newSize/oldSize;
89         int dy2 = src[k][3]*newSize/oldSize;
90         dst[k].p0 = dy1*widthStep + dx1;
91         dst[k].p1 = dy2*widthStep + dx1;
92         dst[k].p2 = dy1*widthStep + dx2;
93         dst[k].p3 = dy2*widthStep + dx2;
94         dst[k].w = src[k][4]/((float)(dx2-dx1)*(dy2-dy1));
95     }
96 }
97 
icvFastHessianDetector(const CvMat * sum,const CvMat * mask_sum,CvMemStorage * storage,const CvSURFParams * params)98 static CvSeq* icvFastHessianDetector( const CvMat* sum, const CvMat* mask_sum,
99     CvMemStorage* storage, const CvSURFParams* params )
100 {
101     CvSeq* points = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvSURFPoint), storage );
102 
103     int totalLayers = params->nOctaves*(params->nOctaveLayers+2);
104     CvMat** hessians = (CvMat**)cvStackAlloc(totalLayers*sizeof(hessians[0]));
105     CvMat** traces = (CvMat**)cvStackAlloc(totalLayers*sizeof(traces[0]));
106     int size, *sizeCache = (int*)cvStackAlloc(totalLayers*sizeof(sizeCache[0]));
107     int scale, *scaleCache = (int*)cvStackAlloc(totalLayers*sizeof(scaleCache[0]));
108 
109     const int NX=3, NY=3, NXY=4, SIZE0=9;
110     int dx_s[NX][5] = { {0, 2, 3, 7, 1}, {3, 2, 6, 7, -2}, {6, 2, 9, 7, 1} };
111     int dy_s[NY][5] = { {2, 0, 7, 3, 1}, {2, 3, 7, 6, -2}, {2, 6, 7, 9, 1} };
112     int dxy_s[NXY][5] = { {1, 1, 4, 4, 1}, {5, 1, 8, 4, -1}, {1, 5, 4, 8, -1}, {5, 5, 8, 8, 1} };
113     int dm[1][5] = { {0, 0, 9, 9, 1} };
114     CvSurfHF Dx[NX], Dy[NY], Dxy[NXY], Dm;
115     double dx = 0, dy = 0, dxy = 0;
116     int hessian_rows, hessian_cols;
117 
118     int octave, sc;
119     int i, j, k, z;
120     int* xofs = (int*)cvStackAlloc(sum->cols*sizeof(xofs[0]));
121 
122     /* hessian detector */
123     for( octave = k = 0; octave < params->nOctaves; octave++ )
124     {
125         for( sc = -1; sc <= params->nOctaveLayers; sc++, k++ )
126         {
127             if ( sc < 0 )
128                 sizeCache[k] = size = 7 << octave; // gaussian scale 1.0;
129             else
130                 sizeCache[k] = size = (sc*6 + 9) << octave; // gaussian scale size*1.2/9.;
131             scaleCache[k] = scale = MAX(size, SIZE0);
132 
133             hessian_rows = (sum->rows)*SIZE0/scale;
134             hessian_cols = (sum->cols)*SIZE0/scale;
135             hessians[k] = cvCreateMat( hessian_rows, hessian_cols, CV_32FC1 );
136             traces[k] = cvCreateMat( hessian_rows, hessian_cols, CV_32FC1 );
137 
138             icvResizeHaarPattern( dx_s, Dx, NX, SIZE0, size, sum->cols );
139             icvResizeHaarPattern( dy_s, Dy, NY, SIZE0, size, sum->cols );
140             icvResizeHaarPattern( dxy_s, Dxy, NXY, SIZE0, size, sum->cols );
141             for( i = 0; i < NXY; i++ )
142                 Dxy[i].w *= 0.9f;
143 
144             float* hessian = hessians[k]->data.fl;
145             float* trace = traces[k]->data.fl;
146 
147             for( i = 0; i < hessian_cols*(SIZE0/2); i++ )
148                 hessian[i] = hessian[hessian_cols*hessian_rows-1-i] =
149                 trace[i] = trace[hessian_cols*hessian_rows-1-i] = 0.f;
150 
151             hessian += (SIZE0/2)*(hessian_cols + 1);
152             trace += (SIZE0/2)*(hessian_cols + 1);
153 
154             for( j = 0; j <= hessian_cols - SIZE0; j++ )
155                 xofs[j] = j*scale/SIZE0;
156 
157             for( i = 0; i < hessian_rows - SIZE0; i++,
158                 trace += hessian_cols, hessian += hessian_cols )
159             {
160                 const int* sum_ptr = sum->data.i + sum->cols*(i*scale/SIZE0);
161                 for( j = 0; j < SIZE0/2; j++ )
162                     hessian[-j-1] = hessian[hessian_cols - SIZE0 + j] =
163                     trace[-j-1] = trace[hessian_cols - SIZE0 + j] = 0.f;
164                 for( j = 0; j <= hessian_cols - SIZE0; j++ )
165                 {
166                     const int* s = sum_ptr + xofs[j];
167                     dx = (s[Dx[0].p0] + s[Dx[0].p3] - s[Dx[0].p1] - s[Dx[0].p2])*Dx[0].w +
168                         (s[Dx[1].p0] + s[Dx[1].p3] - s[Dx[1].p1] - s[Dx[1].p2])*Dx[1].w +
169                         (s[Dx[2].p0] + s[Dx[2].p3] - s[Dx[2].p1] - s[Dx[2].p2])*Dx[2].w;
170                     dy = (s[Dy[0].p0] + s[Dy[0].p3] - s[Dy[0].p1] - s[Dy[0].p2])*Dy[0].w +
171                         (s[Dy[1].p0] + s[Dy[1].p3] - s[Dy[1].p1] - s[Dy[1].p2])*Dy[1].w +
172                         (s[Dy[2].p0] + s[Dy[2].p3] - s[Dy[2].p1] - s[Dy[2].p2])*Dy[2].w;
173                     dxy = (s[Dxy[0].p0] + s[Dxy[0].p3] - s[Dxy[0].p1] - s[Dxy[0].p2])*Dxy[0].w +
174                         (s[Dxy[1].p0] + s[Dxy[1].p3] - s[Dxy[1].p1] - s[Dxy[1].p2])*Dxy[1].w +
175                         (s[Dxy[2].p0] + s[Dxy[2].p3] - s[Dxy[2].p1] - s[Dxy[2].p2])*Dxy[2].w +
176                         (s[Dxy[3].p0] + s[Dxy[3].p3] - s[Dxy[3].p1] - s[Dxy[3].p2])*Dxy[3].w;
177                     hessian[j] = (float)(dx*dy - dxy*dxy);
178                     trace[j] = (float)(dx + dy);
179                 }
180             }
181         }
182     }
183 
184     for( octave = 0, k = 1; octave < params->nOctaves; octave++, k+=2 )
185     {
186         for( sc = 0; sc < params->nOctaveLayers; sc++, k++ )
187         {
188             size = sizeCache[k];
189             scale = scaleCache[k];
190             hessian_rows = hessians[k]->rows;
191             hessian_cols = hessians[k]->cols;
192             icvResizeHaarPattern( dm, &Dm, 1, SIZE0, size, mask_sum ? mask_sum->cols : sum->cols );
193             int margin = 5*scaleCache[k+1]/scale;
194             for( i = margin; i < hessian_rows-margin; i++ )
195             {
196                 const float* hessian = hessians[k]->data.fl + i*hessian_cols;
197                 const float* trace = traces[k]->data.fl + i*hessian_cols;
198                 for( j = margin; j < hessian_cols-margin; j++ )
199                 {
200                     float val0 = hessian[j];
201                     if( val0 > params->hessianThreshold )
202                     {
203                         bool suppressed = false;
204                         if( mask_sum )
205                         {
206                             const int* mask_ptr = mask_sum->data.i +
207                                 mask_sum->cols*((i-SIZE0/2)*scale/SIZE0) +
208                                 (j - SIZE0/2)*scale/SIZE0;
209                             float mval = icvCalcHaarPattern( mask_ptr, &Dm, 1 );
210                             if( mval < 0.5 )
211                                 continue;
212                         }
213 
214                         /* non-maxima suppression */
215                         for( z = k-1; z < k+2; z++ )
216                         {
217                             int hcols_z = hessians[z]->cols;
218                             const float* hessian = hessians[z]->data.fl + (j*scale+scaleCache[z]/2)/scaleCache[z]-1 +
219                                 ((i*scale + scaleCache[z]/2)/scaleCache[z]-1)*hcols_z;
220                             if( val0 < hessian[0] || val0 < hessian[1] || val0 < hessian[2] ||
221                                 val0 < hessian[hcols_z] || val0 < hessian[hcols_z+1] ||
222                                 val0 < hessian[hcols_z+2] || val0 < hessian[hcols_z*2] ||
223                                 val0 < hessian[hcols_z*2+1] || val0 < hessian[hcols_z*2+2] )
224                             {
225                                 suppressed = true;
226                                 break;
227                             }
228                         }
229                         if( !suppressed )
230                         {
231                             double trace_val = trace[j];
232                             CvSURFPoint point = cvSURFPoint( cvPoint2D32f(j*scale/9.f, i*scale/9.f),
233                                 CV_SIGN(trace_val), sizeCache[k], 0, val0 );
234                             cvSeqPush( points, &point );
235                         }
236                     }
237                 }
238             }
239         }
240     }
241 
242     for( octave = k = 0; octave < params->nOctaves; octave++ )
243         for( sc = -1; sc <= params->nOctaveLayers; sc++, k++ )
244         {
245             cvReleaseMat( &hessians[k] );
246             cvReleaseMat( &traces[k] );
247         }
248     return points;
249 }
250 
251 
252 CV_IMPL void
cvExtractSURF(const CvArr * _img,const CvArr * _mask,CvSeq ** _keypoints,CvSeq ** _descriptors,CvMemStorage * storage,CvSURFParams params)253 cvExtractSURF( const CvArr* _img, const CvArr* _mask,
254                CvSeq** _keypoints, CvSeq** _descriptors,
255                CvMemStorage* storage, CvSURFParams params )
256 {
257     CvMat *sum = 0, *mask1 = 0, *mask_sum = 0;
258 
259     if( _keypoints )
260         *_keypoints = 0;
261     if( _descriptors )
262         *_descriptors = 0;
263 
264     CV_FUNCNAME( "cvExtractSURF" );
265 
266     __BEGIN__;
267 
268     CvSeq *keypoints, *descriptors = 0;
269     CvMat imghdr, *img = cvGetMat(_img, &imghdr);
270     CvMat maskhdr, *mask = _mask ? cvGetMat(_mask, &maskhdr) : 0;
271 
272     int descriptor_size = params.extended ? 128 : 64;
273     const int descriptor_data_type = CV_32F;
274     const int NX=2, NY=2;
275     const float sqrt_2 = 1.4142135623730950488016887242097f;
276     const int PATCH_SZ = 20;
277     const int RS_PATCH_SZ = 30; // ceil((PATCH_SZ+1)*sqrt_2);
278     int dx_s[NX][5] = {{0, 0, 2, 4, -1}, {2, 0, 4, 4, 1}};
279     int dy_s[NY][5] = {{0, 0, 4, 2, 1}, {0, 2, 4, 4, -1}};
280     float G[9] = {0,0,0,0,0,0,0,0,0};
281     CvMat _G = cvMat(1, 9, CV_32F, G);
282     float DW[PATCH_SZ][PATCH_SZ];
283     CvMat _DW = cvMat(PATCH_SZ, PATCH_SZ, CV_32F, DW);
284     CvPoint apt[81];
285     int i, j, k, nangle0 = 0, N;
286 
287     CV_ASSERT( img != 0 && CV_MAT_TYPE(img->type) == CV_8UC1 &&
288         (mask == 0 || (CV_ARE_SIZES_EQ(img,mask) &&
289         CV_MAT_TYPE(mask->type) == CV_8UC1)) &&
290         storage != 0 && params.hessianThreshold >= 0 &&
291         params.nOctaves > 0 && params.nOctaveLayers > 0 );
292 
293     sum = cvCreateMat( img->height+1, img->width+1, CV_32SC1 );
294     cvIntegral( img, sum );
295     if( mask )
296     {
297         mask1 = cvCreateMat( img->height, img->width, CV_8UC1 );
298         mask_sum = cvCreateMat( img->height+1, img->width+1, CV_32SC1 );
299         cvMinS( mask, 1, mask1 );
300         cvIntegral( mask1, mask_sum );
301     }
302     keypoints = icvFastHessianDetector( sum, mask_sum, storage, &params );
303     N = keypoints->total;
304     if( _descriptors )
305     {
306         descriptors = cvCreateSeq( 0, sizeof(CvSeq),
307             descriptor_size*CV_ELEM_SIZE(descriptor_data_type), storage );
308         cvSeqPushMulti( descriptors, 0, N );
309     }
310 
311     CvSepFilter::init_gaussian_kernel( &_G, 2.5 );
312 
313     {
314     const double sigma = 3.3;
315     double c2 = 1./(sigma*sigma*2), gs = 0;
316     for( i = 0; i < PATCH_SZ; i++ )
317     {
318         for( j = 0; j < PATCH_SZ; j++ )
319         {
320             double x = j - PATCH_SZ*0.5, y = i - PATCH_SZ*0.5;
321             double val = exp(-(x*x+y*y)*c2);
322             DW[i][j] = (float)val;
323             gs += val;
324         }
325     }
326     cvScale( &_DW, &_DW, 1./gs );
327     }
328 
329     for( i = -4; i <= 4; i++ )
330         for( j = -4; j <= 4; j++ )
331         {
332             if( i*i + j*j <= 16 )
333                 apt[nangle0++] = cvPoint(j,i);
334         }
335 
336     {
337 #ifdef _OPENMP
338     int nthreads = cvGetNumThreads();
339 #pragma omp parallel for num_threads(nthreads) schedule(dynamic)
340 #endif
341     for( k = 0; k < N; k++ )
342     {
343         const int* sum_ptr = sum->data.i;
344         int sum_cols = sum->cols;
345         int i, j, kk, x, y, nangle;
346         CvSurfHF dx_t[NX], dy_t[NY];
347         float X[81], Y[81], angle[81];
348         uchar PATCH[PATCH_SZ+1][PATCH_SZ+1], RS_PATCH[RS_PATCH_SZ][RS_PATCH_SZ];
349         float DX[PATCH_SZ][PATCH_SZ], DY[PATCH_SZ][PATCH_SZ];
350         CvMat _X = cvMat(1, 81, CV_32F, X);
351         CvMat _Y = cvMat(1, 81, CV_32F, Y);
352         CvMat _angle = cvMat(1, 81, CV_32F, angle);
353         CvMat _patch = cvMat(PATCH_SZ+1, PATCH_SZ+1, CV_8U, PATCH);
354         CvMat _rs_patch = cvMat(RS_PATCH_SZ, RS_PATCH_SZ, CV_8U, RS_PATCH);
355         CvMat _src, *src = img;
356 
357         CvSURFPoint* kp = (CvSURFPoint*)cvGetSeqElem( keypoints, k );
358         CvPoint2D32f center = kp->pt;
359         int size = kp->size;
360         icvResizeHaarPattern( dx_s, dx_t, NX, 9, size, sum->cols );
361         icvResizeHaarPattern( dy_s, dy_t, NY, 9, size, sum->cols );
362         CvPoint pt = cvPointFrom32f(center);
363         float* vec;
364         float alpha0, beta0, sz0, scale0;
365 
366         for( kk = 0, nangle = 0; kk < nangle0; kk++ )
367         {
368             j = apt[kk].x; i = apt[kk].y;
369             int x = pt.x + (j-2)*size/9;
370             int y = pt.y + (i-2)*size/9;
371             const int* ptr;
372             float vx, vy, w;
373             if( (unsigned)y >= (unsigned)sum->rows - size ||
374                 (unsigned)x >= (unsigned)sum->cols - size )
375                 continue;
376             ptr = sum_ptr + x + y*sum_cols;
377             w = G[i+4]*G[j+4];
378             vx = icvCalcHaarPattern( ptr, dx_t, NX )*w;
379             vy = icvCalcHaarPattern( ptr, dy_t, NX )*w;
380             X[nangle] = vx; Y[nangle] = vy;
381             nangle++;
382         }
383         _X.cols = _Y.cols = _angle.cols = nangle;
384         cvCartToPolar( &_X, &_Y, 0, &_angle, 1 );
385 
386         float bestx = 0, besty = 0, descriptor_mod = 0;
387         for( i = 0; i < 360; i += 5 )
388         {
389             float sumx = 0, sumy = 0, temp_mod;
390             for( j = 0; j < nangle; j++ )
391             {
392                 int d = abs(cvRound(angle[j]) - i);
393                 if( d < 60 || d > 300 )
394                 {
395                     sumx += X[j];
396                     sumy += Y[j];
397                 }
398             }
399             temp_mod = sumx*sumx + sumy*sumy;
400             if( temp_mod > descriptor_mod )
401             {
402                 descriptor_mod = temp_mod;
403                 bestx = sumx;
404                 besty = sumy;
405             }
406         }
407 
408         float descriptor_dir = cvFastArctan( besty, bestx );
409         kp->dir = descriptor_dir;
410 
411         if( !_descriptors )
412             continue;
413         descriptor_dir *= (float)(CV_PI/180);
414 
415         alpha0 = (float)cos(descriptor_dir);
416         beta0 = (float)sin(descriptor_dir);
417         sz0 = (float)((PATCH_SZ+1)*size*1.2/9.);
418         scale0 = sz0/(PATCH_SZ+1);
419 
420         if( sz0 > (PATCH_SZ+1)*1.5f )
421         {
422             float rd = (float)(sz0*sqrt_2*0.5);
423             float alpha1 = (alpha0 - beta0)*sqrt_2*0.5f, beta1 = (alpha0 + beta0)*sqrt_2*0.5f;
424             CvRect patch_rect0 = { INT_MAX, INT_MAX, INT_MIN, INT_MIN }, patch_rect, sr_patch_rect;
425 
426             for( i = 0; i < 4; i++ )
427             {
428                 float a, b, r = i < 2 ? rd : -rd;
429                 if( i % 2 == 0 )
430                     a = alpha1, b = beta1;
431                 else
432                     a = -beta1, b = alpha1;
433                 float xf = center.x + r*a;
434                 float yf = center.y - r*b;
435                 x = cvFloor(xf); patch_rect0.x = MIN(patch_rect0.x, x);
436                 y = cvFloor(yf); patch_rect0.y = MIN(patch_rect0.y, y);
437                 x = cvCeil(xf)+1; patch_rect0.width = MAX(patch_rect0.width, x);
438                 y = cvCeil(yf)+1; patch_rect0.height = MAX(patch_rect0.height, y);
439             }
440 
441             patch_rect = patch_rect0;
442             patch_rect.x = MAX(patch_rect.x, 0);
443             patch_rect.y = MAX(patch_rect.y, 0);
444             patch_rect.width = MIN(patch_rect.width, img->width) - patch_rect.x;
445             patch_rect.height = MIN(patch_rect.height, img->height) - patch_rect.y;
446             patch_rect0.width -= patch_rect0.x;
447             patch_rect0.height -= patch_rect0.y;
448 
449             CvMat _src0;
450             float scale = MIN(1.f,MIN((float)RS_PATCH_SZ/patch_rect0.width,
451                 (float)RS_PATCH_SZ/patch_rect0.height));
452             cvGetSubArr( img, &_src0, patch_rect );
453             sr_patch_rect = cvRect(0,0, RS_PATCH_SZ, RS_PATCH_SZ);
454             sr_patch_rect.width = cvRound(patch_rect.width*scale);
455             sr_patch_rect.height = cvRound(patch_rect.height*scale);
456             src = cvGetSubArr( &_rs_patch, &_src, sr_patch_rect );
457             cvResize( &_src0, &_src, CV_INTER_AREA );
458             center.x = RS_PATCH_SZ*0.5f - (patch_rect.x - patch_rect0.x)*scale;
459             center.y = RS_PATCH_SZ*0.5f - (patch_rect.y - patch_rect0.y)*scale;
460             scale0 *= scale;
461         }
462 
463         {
464         float w[] =
465         {
466             alpha0*scale0, beta0*scale0, center.x,
467             -beta0*scale0, alpha0*scale0, center.y
468         };
469         CvMat W = cvMat(2, 3, CV_32F, w);
470         cvGetQuadrangleSubPix( src, &_patch, &W );
471         }
472 
473         for( i = 0; i < PATCH_SZ; i++ )
474             for( j = 0; j < PATCH_SZ; j++ )
475             {
476                 float dw = DW[i][j];
477                 float vx = (PATCH[i][j+1] - PATCH[i][j] + PATCH[i+1][j+1] - PATCH[i+1][j])*dw;
478                 float vy = (PATCH[i+1][j] - PATCH[i][j] + PATCH[i+1][j+1] - PATCH[i][j+1])*dw;
479                 DX[i][j] = vx;
480                 DY[i][j] = vy;
481             }
482 
483         vec = (float*)cvGetSeqElem( descriptors, k );
484         for( kk = 0; kk < (int)(descriptors->elem_size/sizeof(vec[0])); kk++ )
485             vec[kk] = 0;
486         if( params.extended )
487         {
488             /* 128-bin descriptor */
489             for( i = 0; i < 4; i++ )
490                 for( j = 0; j < 4; j++ )
491                 {
492                     for( y = i*5; y < i*5+5; y++ )
493                     {
494                         for( x = j*5; x < j*5+5; x++ )
495                         {
496                             float tx = DX[y][x], ty = DY[y][x];
497                             if( ty >= 0 )
498                             {
499                                 vec[0] += tx;
500                                 vec[1] += (float)fabs(tx);
501                             } else {
502                                 vec[2] += tx;
503                                 vec[3] += (float)fabs(tx);
504                             }
505                             if ( tx >= 0 )
506                             {
507                                 vec[4] += ty;
508                                 vec[5] += (float)fabs(ty);
509                             } else {
510                                 vec[6] += ty;
511                                 vec[7] += (float)fabs(ty);
512                             }
513                         }
514                     }
515                     /* unit vector is essential for contrast invariance */
516                     double normalize = 0;
517                     for( kk = 0; kk < 8; kk++ )
518                         normalize += vec[kk]*vec[kk];
519                     normalize = 1./(sqrt(normalize) + DBL_EPSILON);
520                     for( kk = 0; kk < 8; kk++ )
521                         vec[kk] = (float)(vec[kk]*normalize);
522                     vec += 8;
523                 }
524         }
525         else
526         {
527             /* 64-bin descriptor */
528             for( i = 0; i < 4; i++ )
529                 for( j = 0; j < 4; j++ )
530                 {
531                     for( y = i*5; y < i*5+5; y++ )
532                     {
533                         for( x = j*5; x < j*5+5; x++ )
534                         {
535                             float tx = DX[y][x], ty = DY[y][x];
536                             vec[0] += tx; vec[1] += ty;
537                             vec[2] += (float)fabs(tx); vec[3] += (float)fabs(ty);
538                         }
539                     }
540                     double normalize = 0;
541                     for( kk = 0; kk < 4; kk++ )
542                         normalize += vec[kk]*vec[kk];
543                     normalize = 1./(sqrt(normalize) + DBL_EPSILON);
544                     for( kk = 0; kk < 4; kk++ )
545                         vec[kk] = (float)(vec[kk]*normalize);
546                     vec+=4;
547                 }
548         }
549     }
550     }
551 
552     if( _keypoints )
553         *_keypoints = keypoints;
554     if( _descriptors )
555         *_descriptors = descriptors;
556 
557     __END__;
558 
559     cvReleaseMat( &sum );
560     cvReleaseMat( &mask1 );
561     cvReleaseMat( &mask_sum );
562 }
563