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) 2000, Intel Corporation, 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/****************************************************************************************\
43*    Very fast SAD-based (Sum-of-Absolute-Diffrences) stereo correspondence algorithm.   *
44*    Contributed by Kurt Konolige                                                        *
45\****************************************************************************************/
46
47#include "_cv.h"
48/*
49#undef CV_SSE2
50#define CV_SSE2 1
51#include "emmintrin.h"
52*/
53
54CV_IMPL CvStereoBMState*
55cvCreateStereoBMState( int /*preset*/, int numberOfDisparities )
56{
57    CvStereoBMState* state = 0;
58
59    //CV_FUNCNAME( "cvCreateStereoBMState" );
60
61    __BEGIN__;
62
63    state = (CvStereoBMState*)cvAlloc( sizeof(*state) );
64    if( !state )
65        EXIT;
66
67    state->preFilterType = CV_STEREO_BM_NORMALIZED_RESPONSE;
68    state->preFilterSize = 9;
69    state->preFilterCap = 31;
70    state->SADWindowSize = 15;
71    state->minDisparity = 0;
72    state->numberOfDisparities = numberOfDisparities > 0 ? numberOfDisparities : 64;
73    state->textureThreshold = 10;
74    state->uniquenessRatio = 15;
75    state->speckleRange = state->speckleWindowSize = 0;
76
77    state->preFilteredImg0 = state->preFilteredImg1 = state->slidingSumBuf = 0;
78
79    __END__;
80
81    if( cvGetErrStatus() < 0 )
82        cvReleaseStereoBMState( &state );
83    return state;
84}
85
86
87CV_IMPL void
88cvReleaseStereoBMState( CvStereoBMState** state )
89{
90    CV_FUNCNAME( "cvReleaseStereoBMState" );
91
92    __BEGIN__;
93
94    if( !state )
95        CV_ERROR( CV_StsNullPtr, "" );
96
97    if( !*state )
98        EXIT;
99
100    cvReleaseMat( &(*state)->preFilteredImg0 );
101    cvReleaseMat( &(*state)->preFilteredImg1 );
102    cvReleaseMat( &(*state)->slidingSumBuf );
103    cvFree( state );
104
105    __END__;
106}
107
108static void icvPrefilter( const CvMat* src, CvMat* dst, int winsize, int ftzero, uchar* buf )
109{
110    int x, y, wsz2 = winsize/2;
111    int* vsum = (int*)cvAlignPtr(buf + (wsz2 + 1)*sizeof(vsum[0]), 32);
112    int scale_g = winsize*winsize/8, scale_s = (1024 + scale_g)/(scale_g*2);
113    const int OFS = 256*5, TABSZ = OFS*2 + 256;
114    uchar tab[TABSZ];
115    const uchar* sptr = src->data.ptr;
116    int srcstep = src->step;
117    CvSize size = cvGetMatSize(src);
118
119    scale_g *= scale_s;
120
121    for( x = 0; x < TABSZ; x++ )
122        tab[x] = (uchar)(x - OFS < -ftzero ? 0 : x - OFS > ftzero ? ftzero*2 : x - OFS + ftzero);
123
124    for( x = 0; x < size.width; x++ )
125        vsum[x] = (ushort)(sptr[x]*(wsz2 + 2));
126
127    for( y = 1; y < wsz2; y++ )
128    {
129        for( x = 0; x < size.width; x++ )
130            vsum[x] = (ushort)(vsum[x] + sptr[srcstep*y + x]);
131    }
132
133    for( y = 0; y < size.height; y++ )
134    {
135        const uchar* top = sptr + srcstep*MAX(y-wsz2-1,0);
136        const uchar* bottom = sptr + srcstep*MIN(y+wsz2,size.height-1);
137        const uchar* prev = sptr + srcstep*MAX(y-1,0);
138        const uchar* curr = sptr + srcstep*y;
139        const uchar* next = sptr + srcstep*MIN(y+1,size.height-1);
140        uchar* dptr = dst->data.ptr + dst->step*y;
141        x = 0;
142
143        for( ; x < size.width; x++ )
144            vsum[x] = (ushort)(vsum[x] + bottom[x] - top[x]);
145
146        for( x = 0; x <= wsz2; x++ )
147        {
148            vsum[-x-1] = vsum[0];
149            vsum[size.width+x] = vsum[size.width-1];
150        }
151
152        int sum = vsum[0]*(wsz2 + 1);
153        for( x = 1; x <= wsz2; x++ )
154            sum += vsum[x];
155
156        int val = ((curr[0]*5 + curr[1] + prev[0] + next[0])*scale_g - sum*scale_s) >> 10;
157        dptr[0] = tab[val + OFS];
158
159        for( x = 1; x < size.width-1; x++ )
160        {
161            sum += vsum[x+wsz2] - vsum[x-wsz2-1];
162            val = ((curr[x]*4 + curr[x-1] + curr[x+1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
163            dptr[x] = tab[val + OFS];
164        }
165
166        sum += vsum[x+wsz2] - vsum[x-wsz2-1];
167        val = ((curr[x]*5 + curr[x-1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
168        dptr[x] = tab[val + OFS];
169    }
170}
171
172
173static const int DISPARITY_SHIFT = 4;
174
175#if CV_SSE2
176static void
177icvFindStereoCorrespondenceBM_SSE2( const CvMat* left, const CvMat* right,
178                                    CvMat* disp, CvStereoBMState* state,
179                                    uchar* buf, int _dy0, int _dy1 )
180{
181    int x, y, d;
182    int wsz = state->SADWindowSize, wsz2 = wsz/2;
183    int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
184    int ndisp = state->numberOfDisparities;
185    int mindisp = state->minDisparity;
186    int lofs = MAX(ndisp - 1 + mindisp, 0);
187    int rofs = -MIN(ndisp - 1 + mindisp, 0);
188    int width = left->cols, height = left->rows;
189    int width1 = width - rofs - ndisp + 1;
190    int ftzero = state->preFilterCap;
191    int textureThreshold = state->textureThreshold;
192    int uniquenessRatio = state->uniquenessRatio;
193    short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
194
195    ushort *sad, *hsad0, *hsad, *hsad_sub;
196    int *htext;
197    uchar *cbuf0, *cbuf;
198    const uchar* lptr0 = left->data.ptr + lofs;
199    const uchar* rptr0 = right->data.ptr + rofs;
200    const uchar *lptr, *lptr_sub, *rptr;
201    short* dptr = disp->data.s;
202    int sstep = left->step;
203    int dstep = disp->step/sizeof(dptr[0]);
204    int cstep = (height + dy0 + dy1)*ndisp;
205    const int TABSZ = 256;
206    uchar tab[TABSZ];
207    const __m128i d0_8 = _mm_setr_epi16(0,1,2,3,4,5,6,7), dd_8 = _mm_set1_epi16(8);
208
209    sad = (ushort*)cvAlignPtr(buf + sizeof(sad[0]));
210    hsad0 = (ushort*)cvAlignPtr(sad + ndisp + 1 + dy0*ndisp);
211    htext = (int*)cvAlignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2);
212    cbuf0 = (uchar*)cvAlignPtr(htext + height + wsz2 + 2 + dy0*ndisp);
213
214    for( x = 0; x < TABSZ; x++ )
215        tab[x] = (uchar)abs(x - ftzero);
216
217    // initialize buffers
218    memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
219    memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
220
221    for( x = -wsz2-1; x < wsz2; x++ )
222    {
223        hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
224        lptr = lptr0 + MIN(MAX(x, -lofs), width-lofs-1) - dy0*sstep;
225        rptr = rptr0 + MIN(MAX(x, -rofs), width-rofs-1) - dy0*sstep;
226
227        for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
228        {
229            int lval = lptr[0];
230            for( d = 0; d < ndisp; d++ )
231            {
232                int diff = abs(lval - rptr[d]);
233                cbuf[d] = (uchar)diff;
234                hsad[d] = (ushort)(hsad[d] + diff);
235            }
236            htext[y] += tab[lval];
237        }
238    }
239
240    // initialize the left and right borders of the disparity map
241    for( y = 0; y < height; y++ )
242    {
243        for( x = 0; x < lofs; x++ )
244            dptr[y*dstep + x] = FILTERED;
245        for( x = lofs + width1; x < width; x++ )
246            dptr[y*dstep + x] = FILTERED;
247    }
248    dptr += lofs;
249
250    for( x = 0; x < width1; x++, dptr++ )
251    {
252        int x0 = x - wsz2 - 1, x1 = x + wsz2;
253        const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
254        uchar* cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
255        hsad = hsad0 - dy0*ndisp;
256        lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
257        lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
258        rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
259
260        for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
261             hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
262        {
263            int lval = lptr[0];
264            __m128i lv = _mm_set1_epi8((char)lval), z = _mm_setzero_si128();
265            for( d = 0; d < ndisp; d += 16 )
266            {
267                __m128i rv = _mm_loadu_si128((const __m128i*)(rptr + d));
268                __m128i hsad_l = _mm_load_si128((__m128i*)(hsad + d));
269                __m128i hsad_h = _mm_load_si128((__m128i*)(hsad + d + 8));
270                __m128i cbs = _mm_load_si128((const __m128i*)(cbuf_sub + d));
271                __m128i diff = _mm_adds_epu8(_mm_subs_epu8(lv, rv), _mm_subs_epu8(rv, lv));
272                __m128i diff_h = _mm_sub_epi16(_mm_unpackhi_epi8(diff, z), _mm_unpackhi_epi8(cbs, z));
273                _mm_store_si128((__m128i*)(cbuf + d), diff);
274                diff = _mm_sub_epi16(_mm_unpacklo_epi8(diff, z), _mm_unpacklo_epi8(cbs, z));
275                hsad_h = _mm_add_epi16(hsad_h, diff_h);
276                hsad_l = _mm_add_epi16(hsad_l, diff);
277                _mm_store_si128((__m128i*)(hsad + d), hsad_l);
278                _mm_store_si128((__m128i*)(hsad + d + 8), hsad_h);
279            }
280            htext[y] += tab[lval] - tab[lptr_sub[0]];
281        }
282
283        // fill borders
284        for( y = dy1; y <= wsz2; y++ )
285            htext[height+y] = htext[height+dy1-1];
286        for( y = -wsz2-1; y < -dy0; y++ )
287            htext[y] = htext[-dy0];
288
289        // initialize sums
290        for( d = 0; d < ndisp; d++ )
291            sad[d] = (ushort)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
292
293        hsad = hsad0 + (1 - dy0)*ndisp;
294        for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
295            for( d = 0; d < ndisp; d++ )
296                sad[d] = (ushort)(sad[d] + hsad[d]);
297        int tsum = 0;
298        for( y = -wsz2-1; y < wsz2; y++ )
299            tsum += htext[y];
300
301        // finally, start the real processing
302        for( y = 0; y < height; y++ )
303        {
304            int minsad = INT_MAX, mind = -1;
305            hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
306            hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
307            __m128i minsad8 = _mm_set1_epi16(SHRT_MAX);
308            __m128i mind8 = _mm_set1_epi16(-1), d8 = d0_8, mask;
309
310            for( d = 0; d < ndisp; d += 8 )
311            {
312                __m128i v0 = _mm_load_si128((__m128i*)(hsad_sub + d));
313                __m128i v1 = _mm_load_si128((__m128i*)(hsad + d));
314                __m128i sad8 = _mm_load_si128((__m128i*)(sad + d));
315                sad8 = _mm_sub_epi16(sad8, v0);
316                sad8 = _mm_add_epi16(sad8, v1);
317
318                mask = _mm_cmpgt_epi16(minsad8, sad8);
319                _mm_store_si128((__m128i*)(sad + d), sad8);
320                minsad8 = _mm_min_epi16(minsad8, sad8);
321                mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(d8,mind8),mask));
322                d8 = _mm_add_epi16(d8, dd_8);
323            }
324
325            __m128i minsad82 = _mm_unpackhi_epi64(minsad8, minsad8);
326            __m128i mind82 = _mm_unpackhi_epi64(mind8, mind8);
327            mask = _mm_cmpgt_epi16(minsad8, minsad82);
328            mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
329            minsad8 = _mm_min_epi16(minsad8, minsad82);
330
331            minsad82 = _mm_shufflelo_epi16(minsad8, _MM_SHUFFLE(3,2,3,2));
332            mind82 = _mm_shufflelo_epi16(mind8, _MM_SHUFFLE(3,2,3,2));
333            mask = _mm_cmpgt_epi16(minsad8, minsad82);
334            mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
335            minsad8 = _mm_min_epi16(minsad8, minsad82);
336
337            minsad82 = _mm_shufflelo_epi16(minsad8, 1);
338            mind82 = _mm_shufflelo_epi16(mind8, 1);
339            mask = _mm_cmpgt_epi16(minsad8, minsad82);
340            mind8 = _mm_xor_si128(mind8,_mm_and_si128(_mm_xor_si128(mind82,mind8),mask));
341            mind = (short)_mm_cvtsi128_si32(mind8);
342            minsad = sad[mind];
343            tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
344            if( tsum < textureThreshold )
345            {
346                dptr[y*dstep] = FILTERED;
347                continue;
348            }
349
350            if( uniquenessRatio > 0 )
351            {
352                int thresh = minsad + (minsad * uniquenessRatio/100);
353                __m128i thresh8 = _mm_set1_epi16((short)(thresh + 1));
354                __m128i d1 = _mm_set1_epi16((short)(mind-1)), d2 = _mm_set1_epi16((short)(mind+1));
355                __m128i d8 = d0_8;
356
357                for( d = 0; d < ndisp; d += 8 )
358                {
359                    __m128i sad8 = _mm_load_si128((__m128i*)(sad + d));
360                    __m128i mask = _mm_cmpgt_epi16( thresh8, sad8 );
361                    mask = _mm_and_si128(mask, _mm_or_si128(_mm_cmpgt_epi16(d1,d8), _mm_cmpgt_epi16(d8,d2)));
362                    if( _mm_movemask_epi8(mask) )
363                        break;
364                    d8 = _mm_add_epi16(d8, dd_8);
365                }
366                if( d < ndisp )
367                {
368                    dptr[y*dstep] = FILTERED;
369                    continue;
370                }
371            }
372
373            {
374            sad[-1] = sad[1];
375            sad[ndisp] = sad[ndisp-2];
376            int p = sad[mind+1], n = sad[mind-1], d = p + n - 2*sad[mind];
377            dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*128/d : 0) + 15) >> 4);
378            }
379        }
380    }
381}
382#endif
383
384static void
385icvFindStereoCorrespondenceBM( const CvMat* left, const CvMat* right,
386                               CvMat* disp, CvStereoBMState* state,
387                               uchar* buf, int _dy0, int _dy1 )
388{
389    int x, y, d;
390    int wsz = state->SADWindowSize, wsz2 = wsz/2;
391    int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
392    int ndisp = state->numberOfDisparities;
393    int mindisp = state->minDisparity;
394    int lofs = MAX(ndisp - 1 + mindisp, 0);
395    int rofs = -MIN(ndisp - 1 + mindisp, 0);
396    int width = left->cols, height = left->rows;
397    int width1 = width - rofs - ndisp + 1;
398    int ftzero = state->preFilterCap;
399    int textureThreshold = state->textureThreshold;
400    int uniquenessRatio = state->uniquenessRatio;
401    short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
402
403    int *sad, *hsad0, *hsad, *hsad_sub, *htext;
404    uchar *cbuf0, *cbuf;
405    const uchar* lptr0 = left->data.ptr + lofs;
406    const uchar* rptr0 = right->data.ptr + rofs;
407    const uchar *lptr, *lptr_sub, *rptr;
408    short* dptr = disp->data.s;
409    int sstep = left->step;
410    int dstep = disp->step/sizeof(dptr[0]);
411    int cstep = (height+dy0+dy1)*ndisp;
412    const int TABSZ = 256;
413    uchar tab[TABSZ];
414
415    sad = (int*)cvAlignPtr(buf + sizeof(sad[0]));
416    hsad0 = (int*)cvAlignPtr(sad + ndisp + 1 + dy0*ndisp);
417    htext = (int*)cvAlignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2);
418    cbuf0 = (uchar*)cvAlignPtr(htext + height + wsz2 + 2 + dy0*ndisp);
419
420    for( x = 0; x < TABSZ; x++ )
421        tab[x] = (uchar)abs(x - ftzero);
422
423    // initialize buffers
424    memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
425    memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
426
427    for( x = -wsz2-1; x < wsz2; x++ )
428    {
429        hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
430        lptr = lptr0 + MIN(MAX(x, -lofs), width-lofs-1) - dy0*sstep;
431        rptr = rptr0 + MIN(MAX(x, -rofs), width-rofs-1) - dy0*sstep;
432
433        for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
434        {
435            int lval = lptr[0];
436            for( d = 0; d < ndisp; d++ )
437            {
438                int diff = abs(lval - rptr[d]);
439                cbuf[d] = (uchar)diff;
440                hsad[d] = (int)(hsad[d] + diff);
441            }
442            htext[y] += tab[lval];
443        }
444    }
445
446    // initialize the left and right borders of the disparity map
447    for( y = 0; y < height; y++ )
448    {
449        for( x = 0; x < lofs; x++ )
450            dptr[y*dstep + x] = FILTERED;
451        for( x = lofs + width1; x < width; x++ )
452            dptr[y*dstep + x] = FILTERED;
453    }
454    dptr += lofs;
455
456    for( x = 0; x < width1; x++, dptr++ )
457    {
458        int x0 = x - wsz2 - 1, x1 = x + wsz2;
459        const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
460        uchar* cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
461        hsad = hsad0 - dy0*ndisp;
462        lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
463        lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
464        rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
465
466        for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
467             hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
468        {
469            int lval = lptr[0];
470            for( d = 0; d < ndisp; d++ )
471            {
472                int diff = abs(lval - rptr[d]);
473                cbuf[d] = (uchar)diff;
474                hsad[d] = hsad[d] + diff - cbuf_sub[d];
475            }
476            htext[y] += tab[lval] - tab[lptr_sub[0]];
477        }
478
479        // fill borders
480        for( y = dy1; y <= wsz2; y++ )
481            htext[height+y] = htext[height+dy1-1];
482        for( y = -wsz2-1; y < -dy0; y++ )
483            htext[y] = htext[-dy0];
484
485        // initialize sums
486        for( d = 0; d < ndisp; d++ )
487            sad[d] = (int)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
488
489        hsad = hsad0 + (1 - dy0)*ndisp;
490        for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
491            for( d = 0; d < ndisp; d++ )
492                sad[d] = (int)(sad[d] + hsad[d]);
493        int tsum = 0;
494        for( y = -wsz2-1; y < wsz2; y++ )
495            tsum += htext[y];
496
497        // finally, start the real processing
498        for( y = 0; y < height; y++ )
499        {
500            int minsad = INT_MAX, mind = -1;
501            hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
502            hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
503
504            for( d = 0; d < ndisp; d++ )
505            {
506                int currsad = sad[d] + hsad[d] - hsad_sub[d];
507                sad[d] = currsad;
508                if( currsad < minsad )
509                {
510                    minsad = currsad;
511                    mind = d;
512                }
513            }
514            tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
515            if( tsum < textureThreshold )
516            {
517                dptr[y*dstep] = FILTERED;
518                continue;
519            }
520
521            if( uniquenessRatio > 0 )
522            {
523                int thresh = minsad + (minsad * uniquenessRatio/100);
524                for( d = 0; d < ndisp; d++ )
525                {
526                    if( sad[d] <= thresh && (d < mind-1 || d > mind+1))
527                        break;
528                }
529                if( d < ndisp )
530                {
531                    dptr[y*dstep] = FILTERED;
532                    continue;
533                }
534            }
535
536            {
537            sad[-1] = sad[1];
538            sad[ndisp] = sad[ndisp-2];
539            int p = sad[mind+1], n = sad[mind-1], d = p + n - 2*sad[mind];
540            dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*128/d : 0) + 15) >> 4);
541            }
542        }
543    }
544}
545
546
547CV_IMPL void
548cvFindStereoCorrespondenceBM( const CvArr* leftarr, const CvArr* rightarr,
549                              CvArr* disparr, CvStereoBMState* state )
550{
551    CV_FUNCNAME( "cvFindStereoCorrespondenceBM" );
552
553    __BEGIN__;
554
555    CvMat lstub, *left0 = cvGetMat( leftarr, &lstub );
556    CvMat rstub, *right0 = cvGetMat( rightarr, &rstub );
557    CvMat left, right;
558    CvMat dstub, *disp = cvGetMat( disparr, &dstub );
559    int bufSize0, bufSize1, bufSize, width, width1, height;
560    int wsz, ndisp, mindisp, lofs, rofs;
561    int i, n = cvGetNumThreads();
562
563    if( !CV_ARE_SIZES_EQ(left0, right0) ||
564        !CV_ARE_SIZES_EQ(disp, left0) )
565        CV_ERROR( CV_StsUnmatchedSizes, "All the images must have the same size" );
566
567    if( CV_MAT_TYPE(left0->type) != CV_8UC1 ||
568        !CV_ARE_TYPES_EQ(left0, right0) ||
569        CV_MAT_TYPE(disp->type) != CV_16SC1 )
570        CV_ERROR( CV_StsUnsupportedFormat,
571        "Both input images must have 8uC1 format and the disparity image must have 16sC1 format" );
572
573    if( !state )
574        CV_ERROR( CV_StsNullPtr, "Stereo BM state is NULL." );
575
576    if( state->preFilterType != CV_STEREO_BM_NORMALIZED_RESPONSE )
577        CV_ERROR( CV_StsOutOfRange, "preFilterType must be =CV_STEREO_BM_NORMALIZED_RESPONSE" );
578
579    if( state->preFilterSize < 5 || state->preFilterSize > 255 || state->preFilterSize % 2 == 0 )
580        CV_ERROR( CV_StsOutOfRange, "preFilterSize must be odd and be within 5..255" );
581
582    if( state->preFilterCap < 1 || state->preFilterCap > 63 )
583        CV_ERROR( CV_StsOutOfRange, "preFilterCap must be within 1..63" );
584
585    if( state->SADWindowSize < 5 || state->SADWindowSize > 255 || state->SADWindowSize % 2 == 0 ||
586        state->SADWindowSize >= MIN(left0->cols, left0->rows) )
587        CV_ERROR( CV_StsOutOfRange, "SADWindowSize must be odd, be within 5..255 and "
588                                    "be not larger than image width or height" );
589
590    if( state->numberOfDisparities <= 0 || state->numberOfDisparities % 16 != 0 )
591        CV_ERROR( CV_StsOutOfRange, "numberOfDisparities must be positive and divisble by 16" );
592    if( state->textureThreshold < 0 )
593        CV_ERROR( CV_StsOutOfRange, "texture threshold must be non-negative" );
594    if( state->uniquenessRatio < 0 )
595        CV_ERROR( CV_StsOutOfRange, "uniqueness ratio must be non-negative" );
596
597    if( !state->preFilteredImg0 ||
598        state->preFilteredImg0->cols*state->preFilteredImg0->rows < left0->cols*left0->rows )
599    {
600        cvReleaseMat( &state->preFilteredImg0 );
601        cvReleaseMat( &state->preFilteredImg1 );
602
603        state->preFilteredImg0 = cvCreateMat( left0->rows, left0->cols, CV_8U );
604        state->preFilteredImg1 = cvCreateMat( left0->rows, left0->cols, CV_8U );
605    }
606    left = cvMat(left0->rows, left0->cols, CV_8U, state->preFilteredImg0->data.ptr);
607    right = cvMat(right0->rows, right0->cols, CV_8U, state->preFilteredImg1->data.ptr);
608
609    mindisp = state->minDisparity;
610    ndisp = state->numberOfDisparities;
611
612    width = left0->cols;
613    height = left0->rows;
614    lofs = MAX(ndisp - 1 + mindisp, 0);
615    rofs = -MIN(ndisp - 1 + mindisp, 0);
616    width1 = width - rofs - ndisp + 1;
617    if( lofs >= width || rofs >= width || width1 < 1 )
618    {
619        int FILTERED = (short)((state->minDisparity - 1) << DISPARITY_SHIFT);
620        cvSet( disp, cvScalarAll(FILTERED) );
621        EXIT;
622    }
623
624    wsz = state->SADWindowSize;
625    bufSize0 = (ndisp + 2)*sizeof(int) + (height+wsz+2)*ndisp*sizeof(int) +
626        (height + wsz + 2)*sizeof(int) + (height+wsz+2)*ndisp*(wsz+1)*sizeof(uchar) + 256;
627    bufSize1 = (width + state->preFilterSize + 2)*sizeof(int) + 256;
628    bufSize = MAX(bufSize0, bufSize1);
629    n = MAX(MIN(height/wsz, n), 1);
630
631    if( !state->slidingSumBuf || state->slidingSumBuf->cols < bufSize*n )
632    {
633        cvReleaseMat( &state->slidingSumBuf );
634        state->slidingSumBuf = cvCreateMat( 1, bufSize*n, CV_8U );
635    }
636
637#ifdef _OPENMP
638#pragma omp parallel sections num_threads(n)
639#endif
640    {
641    #ifdef _OPENMP
642    #pragma omp section
643    #endif
644        icvPrefilter( left0, &left, state->preFilterSize,
645            state->preFilterCap, state->slidingSumBuf->data.ptr );
646    #ifdef _OPENMP
647    #pragma omp section
648    #endif
649        icvPrefilter( right0, &right, state->preFilterSize,
650            state->preFilterCap, state->slidingSumBuf->data.ptr + bufSize1*(n>1) );
651    }
652
653#ifdef _OPENMP
654    #pragma omp parallel for num_threads(n) schedule(static)
655#endif
656    for( i = 0; i < n; i++ )
657    {
658        int thread_id = cvGetThreadNum();
659        CvMat left_i, right_i, disp_i;
660        int row0 = i*left.rows/n, row1 = (i+1)*left.rows/n;
661        cvGetRows( &left, &left_i, row0, row1 );
662        cvGetRows( &right, &right_i, row0, row1 );
663        cvGetRows( disp, &disp_i, row0, row1 );
664    #if CV_SSE2
665        if( state->preFilterCap <= 31 && state->SADWindowSize <= 21 )
666        {
667            icvFindStereoCorrespondenceBM_SSE2( &left_i, &right_i, &disp_i, state,
668                state->slidingSumBuf->data.ptr + thread_id*bufSize0, row0, left.rows-row1 );
669        }
670        else
671    #endif
672        {
673            icvFindStereoCorrespondenceBM( &left_i, &right_i, &disp_i, state,
674                state->slidingSumBuf->data.ptr + thread_id*bufSize0, row0, left.rows-row1 );
675        }
676    }
677
678    __END__;
679}
680
681/* End of file. */
682