1 //M*//////////////////////////////////////////////////////////////////////////////////////
\r
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
\r
5 // By downloading, copying, installing or using the software you agree to this license.
\r
6 // If you do not agree to this license, do not download, install,
\r
7 // copy or use the software.
\r
10 // Intel License Agreement
\r
11 // For Open Source Computer Vision Library
\r
13 // Copyright (C) 2000, Intel Corporation, all rights reserved.
\r
14 // Third party copyrights are property of their respective owners.
\r
16 // Redistribution and use in source and binary forms, with or without modification,
\r
17 // are permitted provided that the following conditions are met:
\r
19 // * Redistribution's of source code must retain the above copyright notice,
\r
20 // this list of conditions and the following disclaimer.
\r
22 // * Redistribution's in binary form must reproduce the above copyright notice,
\r
23 // this list of conditions and the following disclaimer in the documentation
\r
24 // and/or other materials provided with the distribution.
\r
26 // * The name of Intel Corporation may not be used to endorse or promote products
\r
27 // derived from this software without specific prior written permission.
\r
29 // This software is provided by the copyright holders and contributors "as is" and
\r
30 // any express or implied warranties, including, but not limited to, the implied
\r
31 // warranties of merchantability and fitness for a particular purpose are disclaimed.
\r
32 // In no event shall the Intel Corporation or contributors be liable for any direct,
\r
33 // indirect, incidental, special, exemplary, or consequential damages
\r
34 // (including, but not limited to, procurement of substitute goods or services;
\r
35 // loss of use, data, or profits; or business interruption) however caused
\r
36 // and on any theory of liability, whether in contract, strict liability,
\r
37 // or tort (including negligence or otherwise) arising in any way out of
\r
38 // the use of this software, even if advised of the possibility of such damage.
\r
42 /****************************************************************************************\
\r
43 * Very fast SAD-based (Sum-of-Absolute-Diffrences) stereo correspondence algorithm. *
\r
44 * Contributed by Kurt Konolige *
\r
45 \****************************************************************************************/
\r
51 #include "emmintrin.h"
\r
54 CV_IMPL CvStereoBMState*
\r
55 cvCreateStereoBMState( int /*preset*/, int numberOfDisparities )
\r
57 CvStereoBMState* state = 0;
\r
59 //CV_FUNCNAME( "cvCreateStereoBMState" );
\r
63 state = (CvStereoBMState*)cvAlloc( sizeof(*state) );
\r
67 state->preFilterType = CV_STEREO_BM_NORMALIZED_RESPONSE;
\r
68 state->preFilterSize = 9;
\r
69 state->preFilterCap = 31;
\r
70 state->SADWindowSize = 15;
\r
71 state->minDisparity = 0;
\r
72 state->numberOfDisparities = numberOfDisparities > 0 ? numberOfDisparities : 64;
\r
73 state->textureThreshold = 10;
\r
74 state->uniquenessRatio = 15;
\r
75 state->speckleRange = state->speckleWindowSize = 0;
\r
76 state->trySmallerWindows = 0;
\r
78 state->preFilteredImg0 = state->preFilteredImg1 = state->slidingSumBuf = 0;
\r
79 state->dbmin = state->dbmax = 0;
\r
83 if( cvGetErrStatus() < 0 )
\r
84 cvReleaseStereoBMState( &state );
\r
90 cvReleaseStereoBMState( CvStereoBMState** state )
\r
92 CV_FUNCNAME( "cvReleaseStereoBMState" );
\r
97 CV_ERROR( CV_StsNullPtr, "" );
\r
102 cvReleaseMat( &(*state)->preFilteredImg0 );
\r
103 cvReleaseMat( &(*state)->preFilteredImg1 );
\r
104 cvReleaseMat( &(*state)->slidingSumBuf );
\r
105 cvReleaseMat( &(*state)->dbmin );
\r
106 cvReleaseMat( &(*state)->dbmax );
\r
112 static void icvPrefilter( const CvMat* src, CvMat* dst, int winsize, int ftzero, uchar* buf )
\r
114 int x, y, wsz2 = winsize/2;
\r
115 int* vsum = (int*)cvAlignPtr(buf + (wsz2 + 1)*sizeof(vsum[0]), 32);
\r
116 int scale_g = winsize*winsize/8, scale_s = (1024 + scale_g)/(scale_g*2);
\r
117 const int OFS = 256*5, TABSZ = OFS*2 + 256;
\r
119 const uchar* sptr = src->data.ptr;
\r
120 int srcstep = src->step;
\r
121 CvSize size = cvGetMatSize(src);
\r
123 scale_g *= scale_s;
\r
125 for( x = 0; x < TABSZ; x++ )
\r
126 tab[x] = (uchar)(x - OFS < -ftzero ? 0 : x - OFS > ftzero ? ftzero*2 : x - OFS + ftzero);
\r
128 for( x = 0; x < size.width; x++ )
\r
129 vsum[x] = (ushort)(sptr[x]*(wsz2 + 2));
\r
131 for( y = 1; y < wsz2; y++ )
\r
133 for( x = 0; x < size.width; x++ )
\r
134 vsum[x] = (ushort)(vsum[x] + sptr[srcstep*y + x]);
\r
137 for( y = 0; y < size.height; y++ )
\r
139 const uchar* top = sptr + srcstep*MAX(y-wsz2-1,0);
\r
140 const uchar* bottom = sptr + srcstep*MIN(y+wsz2,size.height-1);
\r
141 const uchar* prev = sptr + srcstep*MAX(y-1,0);
\r
142 const uchar* curr = sptr + srcstep*y;
\r
143 const uchar* next = sptr + srcstep*MIN(y+1,size.height-1);
\r
144 uchar* dptr = dst->data.ptr + dst->step*y;
\r
147 for( ; x < size.width; x++ )
\r
148 vsum[x] = (ushort)(vsum[x] + bottom[x] - top[x]);
\r
150 for( x = 0; x <= wsz2; x++ )
\r
152 vsum[-x-1] = vsum[0];
\r
153 vsum[size.width+x] = vsum[size.width-1];
\r
156 int sum = vsum[0]*(wsz2 + 1);
\r
157 for( x = 1; x <= wsz2; x++ )
\r
160 int val = ((curr[0]*5 + curr[1] + prev[0] + next[0])*scale_g - sum*scale_s) >> 10;
\r
161 dptr[0] = tab[val + OFS];
\r
163 for( x = 1; x < size.width-1; x++ )
\r
165 sum += vsum[x+wsz2] - vsum[x-wsz2-1];
\r
166 val = ((curr[x]*4 + curr[x-1] + curr[x+1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
\r
167 dptr[x] = tab[val + OFS];
\r
170 sum += vsum[x+wsz2] - vsum[x-wsz2-1];
\r
171 val = ((curr[x]*5 + curr[x-1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
\r
172 dptr[x] = tab[val + OFS];
\r
177 static const int DISPARITY_SHIFT = 4;
\r
181 icvFindStereoCorrespondenceBM_SSE2( const CvMat* left, const CvMat* right,
\r
182 CvMat* disp, const CvMat* dbmin,
\r
183 const CvMat* dbmax, CvStereoBMState* state,
\r
184 int realSADWin, uchar* buf, int _dy0, int _dy1 )
\r
187 int wsz = realSADWin, wsz2 = wsz/2;
\r
188 int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
\r
189 int ndisp = state->numberOfDisparities;
\r
190 int mindisp = state->minDisparity;
\r
191 int lofs = MAX(ndisp - 1 + mindisp, 0);
\r
192 int rofs = -MIN(ndisp - 1 + mindisp, 0);
\r
193 int width = left->cols, height = left->rows;
\r
194 int width1 = width - rofs - ndisp + 1;
\r
195 int ftzero = state->preFilterCap;
\r
196 int textureThreshold = state->textureThreshold;
\r
197 int uniquenessRatio = state->uniquenessRatio;
\r
198 short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
\r
199 int DELTA = (state->numberOfDisparities - 1 - state->minDisparity) << DISPARITY_SHIFT;
\r
201 ushort *sad, *hsad0, *hsad, *hsad_sub;
\r
203 uchar *cbuf0, *cbuf;
\r
204 const uchar* lptr0 = left->data.ptr + lofs;
\r
205 const uchar* rptr0 = right->data.ptr + rofs;
\r
206 const uchar *lptr, *lptr_sub, *rptr;
\r
207 short* dptr = disp->data.s;
\r
208 int sstep = left->step;
\r
209 int dstep = disp->step/sizeof(dptr[0]);
\r
210 int cstep = (height + dy0 + dy1)*ndisp;
\r
211 const int TABSZ = 256;
\r
213 const __m128i d0_8 = _mm_setr_epi16(0,1,2,3,4,5,6,7), dd_8 = _mm_set1_epi16(8);
\r
215 sad = (ushort*)cvAlignPtr(buf + sizeof(sad[0]));
\r
216 hsad0 = (ushort*)cvAlignPtr(sad + ndisp + 1 + dy0*ndisp);
\r
217 htext = (int*)cvAlignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2);
\r
218 cbuf0 = (uchar*)cvAlignPtr(htext + height + wsz2 + 2 + dy0*ndisp);
\r
220 for( x = 0; x < TABSZ; x++ )
\r
221 tab[x] = (uchar)abs(x - ftzero);
\r
223 // initialize buffers
\r
224 memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
\r
225 memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
\r
227 for( x = -wsz2-1; x < wsz2; x++ )
\r
229 hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
\r
230 lptr = lptr0 + MIN(MAX(x, -lofs), width-lofs-1) - dy0*sstep;
\r
231 rptr = rptr0 + MIN(MAX(x, -rofs), width-rofs-1) - dy0*sstep;
\r
233 for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
\r
235 int lval = lptr[0];
\r
236 for( d = 0; d < ndisp; d++ )
\r
238 int diff = abs(lval - rptr[d]);
\r
239 cbuf[d] = (uchar)diff;
\r
240 hsad[d] = (ushort)(hsad[d] + diff);
\r
242 htext[y] += tab[lval];
\r
246 // initialize the left and right borders of the disparity map
\r
247 for( y = 0; y < height; y++ )
\r
249 for( x = 0; x < lofs; x++ )
\r
250 dptr[y*dstep + x] = FILTERED;
\r
251 for( x = lofs + width1; x < width; x++ )
\r
252 dptr[y*dstep + x] = FILTERED;
\r
256 for( x = 0; x < width1; x++, dptr++ )
\r
258 int x0 = x - wsz2 - 1, x1 = x + wsz2;
\r
259 const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
\r
260 uchar* cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
\r
261 hsad = hsad0 - dy0*ndisp;
\r
262 lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
\r
263 lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
\r
264 rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
\r
266 for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
\r
267 hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
\r
269 int lval = lptr[0];
\r
270 __m128i lv = _mm_set1_epi8((char)lval), z = _mm_setzero_si128();
\r
271 for( d = 0; d < ndisp; d += 16 )
\r
273 __m128i rv = _mm_loadu_si128((const __m128i*)(rptr + d));
\r
274 __m128i hsad_l = _mm_load_si128((__m128i*)(hsad + d));
\r
275 __m128i hsad_h = _mm_load_si128((__m128i*)(hsad + d + 8));
\r
276 __m128i cbs = _mm_load_si128((const __m128i*)(cbuf_sub + d));
\r
277 __m128i diff = _mm_adds_epu8(_mm_subs_epu8(lv, rv), _mm_subs_epu8(rv, lv));
\r
278 __m128i diff_h = _mm_sub_epi16(_mm_unpackhi_epi8(diff, z), _mm_unpackhi_epi8(cbs, z));
\r
279 _mm_store_si128((__m128i*)(cbuf + d), diff);
\r
280 diff = _mm_sub_epi16(_mm_unpacklo_epi8(diff, z), _mm_unpacklo_epi8(cbs, z));
\r
281 hsad_h = _mm_add_epi16(hsad_h, diff_h);
\r
282 hsad_l = _mm_add_epi16(hsad_l, diff);
\r
283 _mm_store_si128((__m128i*)(hsad + d), hsad_l);
\r
284 _mm_store_si128((__m128i*)(hsad + d + 8), hsad_h);
\r
286 htext[y] += tab[lval] - tab[lptr_sub[0]];
\r
290 for( y = dy1; y <= wsz2; y++ )
\r
291 htext[height+y] = htext[height+dy1-1];
\r
292 for( y = -wsz2-1; y < -dy0; y++ )
\r
293 htext[y] = htext[-dy0];
\r
296 for( d = 0; d < ndisp; d++ )
\r
297 sad[d] = (ushort)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
\r
299 hsad = hsad0 + (1 - dy0)*ndisp;
\r
300 for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
\r
301 for( d = 0; d < ndisp; d++ )
\r
302 sad[d] = (ushort)(sad[d] + hsad[d]);
\r
304 for( y = -wsz2-1; y < wsz2; y++ )
\r
307 // finally, start the real processing
\r
308 for( y = 0; y < height; y++ )
\r
310 int minsad = INT_MAX, mind = -1;
\r
311 hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
\r
312 hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
\r
313 __m128i minsad8 = _mm_set1_epi16(SHRT_MAX);
\r
314 __m128i mind8 = _mm_set1_epi16(-1), d8 = d0_8, mask;
\r
316 for( d = 0; d < ndisp; d += 8 )
\r
318 __m128i v0 = _mm_load_si128((__m128i*)(hsad_sub + d));
\r
319 __m128i v1 = _mm_load_si128((__m128i*)(hsad + d));
\r
320 __m128i sad8 = _mm_load_si128((__m128i*)(sad + d));
\r
321 sad8 = _mm_sub_epi16(sad8, v0);
\r
322 sad8 = _mm_add_epi16(sad8, v1);
\r
324 mask = _mm_cmpgt_epi16(minsad8, sad8);
\r
325 _mm_store_si128((__m128i*)(sad + d), sad8);
\r
326 minsad8 = _mm_min_epi16(minsad8, sad8);
\r
327 mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(d8,mind8),mask));
\r
328 d8 = _mm_add_epi16(d8, dd_8);
\r
331 __m128i minsad82 = _mm_unpackhi_epi64(minsad8, minsad8);
\r
332 __m128i mind82 = _mm_unpackhi_epi64(mind8, mind8);
\r
333 mask = _mm_cmpgt_epi16(minsad8, minsad82);
\r
334 mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
\r
335 minsad8 = _mm_min_epi16(minsad8, minsad82);
\r
337 minsad82 = _mm_shufflelo_epi16(minsad8, _MM_SHUFFLE(3,2,3,2));
\r
338 mind82 = _mm_shufflelo_epi16(mind8, _MM_SHUFFLE(3,2,3,2));
\r
339 mask = _mm_cmpgt_epi16(minsad8, minsad82);
\r
340 mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
\r
341 minsad8 = _mm_min_epi16(minsad8, minsad82);
\r
343 minsad82 = _mm_shufflelo_epi16(minsad8, 1);
\r
344 mind82 = _mm_shufflelo_epi16(mind8, 1);
\r
345 mask = _mm_cmpgt_epi16(minsad8, minsad82);
\r
346 mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
\r
347 mind = (short)_mm_cvtsi128_si32(mind8);
\r
348 minsad = sad[mind];
\r
349 tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
\r
350 if( tsum < textureThreshold )
\r
353 dptr[y*dstep] = FILTERED;
\r
357 if( uniquenessRatio > 0 )
\r
359 int thresh = minsad + (minsad * uniquenessRatio/100);
\r
360 __m128i thresh8 = _mm_set1_epi16((short)(thresh + 1));
\r
361 __m128i d1 = _mm_set1_epi16((short)(mind-1)), d2 = _mm_set1_epi16((short)(mind+1));
\r
364 for( d = 0; d < ndisp; d += 8 )
\r
366 __m128i sad8 = _mm_load_si128((__m128i*)(sad + d));
\r
367 __m128i mask = _mm_cmpgt_epi16( thresh8, sad8 );
\r
368 mask = _mm_and_si128(mask, _mm_or_si128(_mm_cmpgt_epi16(d1,d8), _mm_cmpgt_epi16(d8,d2)));
\r
369 if( _mm_movemask_epi8(mask) )
\r
371 d8 = _mm_add_epi16(d8, dd_8);
\r
376 dptr[y*dstep] = FILTERED;
\r
383 int maxval = ((const short*)(dbmin->data.ptr + dbmin->step*y))[x];
\r
384 int minval = ((const short*)(dbmax->data.ptr + dbmax->step*y))[x];
\r
385 minval = (DELTA - minval) >> DISPARITY_SHIFT;
\r
386 maxval = (DELTA - maxval + (1<<DISPARITY_SHIFT)-1) >> DISPARITY_SHIFT;
\r
387 if( d < minval || d > maxval )
\r
393 sad[ndisp] = sad[ndisp-2];
\r
394 int p = sad[mind+1], n = sad[mind-1], d = p + n - 2*sad[mind];
\r
395 dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*128/d : 0) + 15) >> 4);
\r
403 icvFindStereoCorrespondenceBM( const CvMat* left, const CvMat* right,
\r
404 CvMat* disp, const CvMat* dbmin,
\r
405 const CvMat* dbmax, CvStereoBMState* state,
\r
406 int realSADWin, uchar* buf, int _dy0, int _dy1 )
\r
409 int wsz = realSADWin, wsz2 = wsz/2;
\r
410 int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
\r
411 int ndisp = state->numberOfDisparities;
\r
412 int mindisp = state->minDisparity;
\r
413 int lofs = MAX(ndisp - 1 + mindisp, 0);
\r
414 int rofs = -MIN(ndisp - 1 + mindisp, 0);
\r
415 int width = left->cols, height = left->rows;
\r
416 int width1 = width - rofs - ndisp + 1;
\r
417 int ftzero = state->preFilterCap;
\r
418 int textureThreshold = state->textureThreshold;
\r
419 int uniquenessRatio = state->uniquenessRatio;
\r
420 short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
\r
421 int DELTA = (state->numberOfDisparities - 1 - state->minDisparity) << DISPARITY_SHIFT;
\r
423 int *sad, *hsad0, *hsad, *hsad_sub, *htext;
\r
424 uchar *cbuf0, *cbuf;
\r
425 const uchar* lptr0 = left->data.ptr + lofs;
\r
426 const uchar* rptr0 = right->data.ptr + rofs;
\r
427 const uchar *lptr, *lptr_sub, *rptr;
\r
428 short* dptr = disp->data.s;
\r
429 int sstep = left->step;
\r
430 int dstep = disp->step/sizeof(dptr[0]);
\r
431 int cstep = (height+dy0+dy1)*ndisp;
\r
432 const int TABSZ = 256;
\r
435 sad = (int*)cvAlignPtr(buf + sizeof(sad[0]));
\r
436 hsad0 = (int*)cvAlignPtr(sad + ndisp + 1 + dy0*ndisp);
\r
437 htext = (int*)cvAlignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2);
\r
438 cbuf0 = (uchar*)cvAlignPtr(htext + height + wsz2 + 2 + dy0*ndisp);
\r
440 for( x = 0; x < TABSZ; x++ )
\r
441 tab[x] = (uchar)abs(x - ftzero);
\r
443 // initialize buffers
\r
444 memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
\r
445 memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
\r
447 for( x = -wsz2-1; x < wsz2; x++ )
\r
449 hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
\r
450 lptr = lptr0 + MIN(MAX(x, -lofs), width-lofs-1) - dy0*sstep;
\r
451 rptr = rptr0 + MIN(MAX(x, -rofs), width-rofs-1) - dy0*sstep;
\r
453 for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
\r
455 int lval = lptr[0];
\r
456 for( d = 0; d < ndisp; d++ )
\r
458 int diff = abs(lval - rptr[d]);
\r
459 cbuf[d] = (uchar)diff;
\r
460 hsad[d] = (int)(hsad[d] + diff);
\r
462 htext[y] += tab[lval];
\r
466 // initialize the left and right borders of the disparity map
\r
467 for( y = 0; y < height; y++ )
\r
469 for( x = 0; x < lofs; x++ )
\r
470 dptr[y*dstep + x] = FILTERED;
\r
471 for( x = lofs + width1; x < width; x++ )
\r
472 dptr[y*dstep + x] = FILTERED;
\r
476 for( x = 0; x < width1; x++, dptr++ )
\r
478 int x0 = x - wsz2 - 1, x1 = x + wsz2;
\r
479 const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
\r
480 uchar* cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
\r
481 hsad = hsad0 - dy0*ndisp;
\r
482 lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
\r
483 lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
\r
484 rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
\r
486 for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
\r
487 hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
\r
489 int lval = lptr[0];
\r
490 for( d = 0; d < ndisp; d++ )
\r
492 int diff = abs(lval - rptr[d]);
\r
493 cbuf[d] = (uchar)diff;
\r
494 hsad[d] = hsad[d] + diff - cbuf_sub[d];
\r
496 htext[y] += tab[lval] - tab[lptr_sub[0]];
\r
500 for( y = dy1; y <= wsz2; y++ )
\r
501 htext[height+y] = htext[height+dy1-1];
\r
502 for( y = -wsz2-1; y < -dy0; y++ )
\r
503 htext[y] = htext[-dy0];
\r
506 for( d = 0; d < ndisp; d++ )
\r
507 sad[d] = (int)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
\r
509 hsad = hsad0 + (1 - dy0)*ndisp;
\r
510 for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
\r
511 for( d = 0; d < ndisp; d++ )
\r
512 sad[d] = (int)(sad[d] + hsad[d]);
\r
514 for( y = -wsz2-1; y < wsz2; y++ )
\r
517 // finally, start the real processing
\r
518 for( y = 0; y < height; y++ )
\r
520 int minsad = INT_MAX, mind = -1;
\r
521 hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
\r
522 hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
\r
524 for( d = 0; d < ndisp; d++ )
\r
526 int currsad = sad[d] + hsad[d] - hsad_sub[d];
\r
528 if( currsad < minsad )
\r
534 tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
\r
535 if( tsum < textureThreshold )
\r
538 dptr[y*dstep] = FILTERED;
\r
542 if( uniquenessRatio > 0 )
\r
544 int thresh = minsad + (minsad * uniquenessRatio/100);
\r
545 for( d = 0; d < ndisp; d++ )
\r
547 if( sad[d] <= thresh && (d < mind-1 || d > mind+1))
\r
553 dptr[y*dstep] = FILTERED;
\r
560 int maxval = ((const short*)(dbmin->data.ptr + dbmin->step*y))[x];
\r
561 int minval = ((const short*)(dbmax->data.ptr + dbmax->step*y))[x];
\r
562 minval = (DELTA - minval) >> DISPARITY_SHIFT;
\r
563 maxval = (DELTA - maxval + (1<<DISPARITY_SHIFT)-1) >> DISPARITY_SHIFT;
\r
564 if( d < minval || d > maxval )
\r
570 sad[ndisp] = sad[ndisp-2];
\r
571 int p = sad[mind+1], n = sad[mind-1], d = p + n - 2*sad[mind];
\r
572 dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*128/d : 0) + 15) >> 4);
\r
579 cvFindStereoCorrespondenceBM( const CvArr* leftarr, const CvArr* rightarr,
\r
580 CvArr* disparr, CvStereoBMState* state )
\r
582 CV_FUNCNAME( "cvFindStereoCorrespondenceBM" );
\r
586 CvMat lstub, *left0 = cvGetMat( leftarr, &lstub );
\r
587 CvMat rstub, *right0 = cvGetMat( rightarr, &rstub );
\r
589 CvMat dstub, *disp = cvGetMat( disparr, &dstub );
\r
590 int bufSize0, bufSize1, bufSize, width, width1, height;
\r
591 int wsz, ndisp, mindisp, lofs, rofs;
\r
592 int i, n = cvGetNumThreads();
\r
593 int FILTERED, SMALL_WIN_SIZE;
\r
595 if( !CV_ARE_SIZES_EQ(left0, right0) ||
\r
596 !CV_ARE_SIZES_EQ(disp, left0) )
\r
597 CV_ERROR( CV_StsUnmatchedSizes, "All the images must have the same size" );
\r
599 if( CV_MAT_TYPE(left0->type) != CV_8UC1 ||
\r
600 !CV_ARE_TYPES_EQ(left0, right0) ||
\r
601 CV_MAT_TYPE(disp->type) != CV_16SC1 )
\r
602 CV_ERROR( CV_StsUnsupportedFormat,
\r
603 "Both input images must have 8uC1 format and the disparity image must have 16sC1 format" );
\r
606 CV_ERROR( CV_StsNullPtr, "Stereo BM state is NULL." );
\r
608 if( state->preFilterType != CV_STEREO_BM_NORMALIZED_RESPONSE )
\r
609 CV_ERROR( CV_StsOutOfRange, "preFilterType must be =CV_STEREO_BM_NORMALIZED_RESPONSE" );
\r
611 if( state->preFilterSize < 5 || state->preFilterSize > 255 || state->preFilterSize % 2 == 0 )
\r
612 CV_ERROR( CV_StsOutOfRange, "preFilterSize must be odd and be within 5..255" );
\r
614 if( state->preFilterCap < 1 || state->preFilterCap > 63 )
\r
615 CV_ERROR( CV_StsOutOfRange, "preFilterCap must be within 1..63" );
\r
617 if( state->SADWindowSize < 5 || state->SADWindowSize > 255 || state->SADWindowSize % 2 == 0 ||
\r
618 state->SADWindowSize >= MIN(left0->cols, left0->rows) )
\r
619 CV_ERROR( CV_StsOutOfRange, "SADWindowSize must be odd, be within 5..255 and "
\r
620 "be not larger than image width or height" );
\r
622 if( state->numberOfDisparities <= 0 || state->numberOfDisparities % 16 != 0 )
\r
623 CV_ERROR( CV_StsOutOfRange, "numberOfDisparities must be positive and divisble by 16" );
\r
624 if( state->textureThreshold < 0 )
\r
625 CV_ERROR( CV_StsOutOfRange, "texture threshold must be non-negative" );
\r
626 if( state->uniquenessRatio < 0 )
\r
627 CV_ERROR( CV_StsOutOfRange, "uniqueness ratio must be non-negative" );
\r
629 if( !state->preFilteredImg0 ||
\r
630 state->preFilteredImg0->cols*state->preFilteredImg0->rows < left0->cols*left0->rows )
\r
632 cvReleaseMat( &state->preFilteredImg0 );
\r
633 cvReleaseMat( &state->preFilteredImg1 );
\r
635 state->preFilteredImg0 = cvCreateMat( left0->rows, left0->cols, CV_8U );
\r
636 state->preFilteredImg1 = cvCreateMat( left0->rows, left0->cols, CV_8U );
\r
638 left = cvMat(left0->rows, left0->cols, CV_8U, state->preFilteredImg0->data.ptr);
\r
639 right = cvMat(right0->rows, right0->cols, CV_8U, state->preFilteredImg1->data.ptr);
\r
641 mindisp = state->minDisparity;
\r
642 ndisp = state->numberOfDisparities;
\r
644 width = left0->cols;
\r
645 height = left0->rows;
\r
646 lofs = MAX(ndisp - 1 + mindisp, 0);
\r
647 rofs = -MIN(ndisp - 1 + mindisp, 0);
\r
648 width1 = width - rofs - ndisp + 1;
\r
649 FILTERED = (short)((state->minDisparity - 1) << DISPARITY_SHIFT);
\r
651 if( lofs >= width || rofs >= width || width1 < 1 )
\r
653 cvSet( disp, cvScalarAll(FILTERED) );
\r
657 wsz = state->SADWindowSize;
\r
658 bufSize0 = (ndisp + 2)*sizeof(int) + (height+wsz+2)*ndisp*sizeof(int) +
\r
659 (height + wsz + 2)*sizeof(int) + (height+wsz+2)*ndisp*(wsz+1)*sizeof(uchar) + 256;
\r
660 bufSize1 = (width + state->preFilterSize + 2)*sizeof(int) + 256;
\r
661 bufSize = MAX(bufSize0, bufSize1);
\r
662 n = MAX(MIN(height/wsz, n), 1);
\r
664 if( !state->slidingSumBuf || state->slidingSumBuf->cols < bufSize*n )
\r
666 cvReleaseMat( &state->slidingSumBuf );
\r
667 state->slidingSumBuf = cvCreateMat( 1, bufSize*n, CV_8U );
\r
671 #pragma omp parallel sections num_threads(n)
\r
675 #pragma omp section
\r
677 icvPrefilter( left0, &left, state->preFilterSize,
\r
678 state->preFilterCap, state->slidingSumBuf->data.ptr );
\r
680 #pragma omp section
\r
682 icvPrefilter( right0, &right, state->preFilterSize,
\r
683 state->preFilterCap, state->slidingSumBuf->data.ptr + bufSize1*(n>1) );
\r
687 #pragma omp parallel for num_threads(n) schedule(static)
\r
689 for( i = 0; i < n; i++ )
\r
691 int thread_id = cvGetThreadNum();
\r
692 CvMat left_i, right_i, disp_i;
\r
693 int row0 = i*left.rows/n, row1 = (i+1)*left.rows/n;
\r
694 cvGetRows( &left, &left_i, row0, row1 );
\r
695 cvGetRows( &right, &right_i, row0, row1 );
\r
696 cvGetRows( disp, &disp_i, row0, row1 );
\r
698 if( state->preFilterCap <= 31 && state->SADWindowSize <= 21 )
\r
700 icvFindStereoCorrespondenceBM_SSE2( &left_i, &right_i, &disp_i, 0, 0, state,
\r
701 state->SADWindowSize, state->slidingSumBuf->data.ptr + thread_id*bufSize0,
\r
702 row0, left.rows-row1 );
\r
707 icvFindStereoCorrespondenceBM( &left_i, &right_i, &disp_i, 0, 0, state,
\r
708 state->SADWindowSize, state->slidingSumBuf->data.ptr + thread_id*bufSize0,
\r
709 row0, left.rows-row1 );
\r
713 SMALL_WIN_SIZE = MAX((state->SADWindowSize/2)|1, 9);
\r
715 if( !state->trySmallerWindows || SMALL_WIN_SIZE >= state->SADWindowSize )
\r
718 if( !state->dbmin || !CV_ARE_SIZES_EQ(state->dbmin, disp) )
\r
720 cvReleaseMat( &state->dbmin );
\r
721 cvReleaseMat( &state->dbmax );
\r
722 state->dbmin = cvCreateMat( disp->rows, disp->cols, CV_16SC1 );
\r
723 state->dbmax = cvCreateMat( disp->rows, disp->cols, CV_16SC1 );
\r
726 for( i = 0; i < disp->rows; i++ )
\r
729 short* minptr = (short*)(state->dbmin->data.ptr + state->dbmin->step*i);
\r
730 short* maxptr = (short*)(state->dbmax->data.ptr + state->dbmax->step*i);
\r
732 for( j = 0; j < disp->cols; j++ )
\r
734 short dval = ((const short*)(disp->data.ptr + disp->step*i))[j];
\r
735 if( dval < (state->minDisparity << DISPARITY_SHIFT) )
\r
736 minptr[j] = maxptr[j] = dval;
\r
739 minptr[j] = SHRT_MAX;
\r
740 maxptr[j] = SHRT_MIN;
\r
745 cvErode(state->dbmin, state->dbmin, 0, SMALL_WIN_SIZE );
\r
746 cvDilate(state->dbmax, state->dbmax, 0, SMALL_WIN_SIZE );
\r
749 #pragma omp parallel for num_threads(n) schedule(static)
\r
751 for( i = 0; i < n; i++ )
\r
753 int thread_id = cvGetThreadNum();
\r
754 CvMat left_i, right_i, disp_i, dbmin_i, dbmax_i;
\r
755 int row0 = i*left.rows/n, row1 = (i+1)*left.rows/n;
\r
756 cvGetRows( &left, &left_i, row0, row1 );
\r
757 cvGetRows( &right, &right_i, row0, row1 );
\r
758 cvGetRows( disp, &disp_i, row0, row1 );
\r
759 cvGetRows( state->dbmin, &dbmin_i, row0, row1 );
\r
760 cvGetRows( state->dbmax, &dbmax_i, row0, row1 );
\r
762 if( state->preFilterCap <= 31 && SMALL_WIN_SIZE <= 21 )
\r
764 icvFindStereoCorrespondenceBM_SSE2( &left_i, &right_i,
\r
765 &disp_i, &dbmin_i, &dbmax_i, state, SMALL_WIN_SIZE,
\r
766 state->slidingSumBuf->data.ptr + thread_id*bufSize0,
\r
767 row0, left.rows-row1 );
\r
772 icvFindStereoCorrespondenceBM( &left_i, &right_i,
\r
773 &disp_i, &dbmin_i, &dbmax_i, state, SMALL_WIN_SIZE,
\r
774 state->slidingSumBuf->data.ptr + thread_id*bufSize0,
\r
775 row0, left.rows-row1 );
\r
785 StereoBM::StereoBM()
\r
786 { state = cvCreateStereoBMState(); }
\r
788 StereoBM::StereoBM(int _preset, int _ndisparities, int _SADWindowSize)
\r
789 { init(_preset, _ndisparities, _SADWindowSize); }
\r
791 void StereoBM::init(int _preset, int _ndisparities, int _SADWindowSize)
\r
793 state = cvCreateStereoBMState(_preset, _ndisparities);
\r
794 state->SADWindowSize = _SADWindowSize;
\r
797 void StereoBM::operator()( const Mat& left, const Mat& right, Mat& disparity )
\r
799 disparity.create(left.size(), CV_16SC1);
\r
800 CvMat _left = left, _right = right, _disparity = disparity;
\r
801 cvFindStereoCorrespondenceBM(&_left, &_right, &_disparity, state);
\r