OpenCV SGBM算法自己的一些注释，
网上有类似文章，自己再做一些补充，
参考 opencvSGBM半全局立体匹配算法的研究(1)_zhubaohua_bupt的博客-CSDN博客 这篇文章对左右一致性检测写的很详细
OpenCV中SGBM中BT代价计算_Tomorrow_Maple的博客-CSDN博客_bt算法
这篇文章代价计算，看这个就明白了
还有一篇文章，OpenCV源代码分析——SGBM - 知乎
同样参考了李迎松的SGM算法的文章，写得很详细，赞一个
/*
This is a variation of
"Stereo Processing by Semiglobal Matching and Mutual Information"
by Heiko Hirschmuller.
We match blocks rather than individual pixels, thus the algorithm is called
SGBM (Semi-global block matching)
*/
void compute( InputArray leftarr, InputArray rightarr, OutputArray disparr )
{
CV_INSTRUMENT_REGION()
Mat left = leftarr.getMat(), right = rightarr.getMat();
CV_Assert( left.size() == right.size() && left.type() == right.type() &&
left.depth() == CV_8U );
disparr.create( left.size(), CV_16S );
Mat disp = disparr.getMat();
if(params.mode==MODE_SGBM_3WAY)
computeDisparity3WaySGBM( left, right, disp, params, buffers, num_stripes );
else if(params.mode==MODE_HH4)
computeDisparitySGBM_HH4( left, right, disp, params, buffer );
else
computeDisparitySGBM( left, right, disp, params, buffer );
先SGBM，再medianBlur，再去最小连通域
medianBlur(disp, disp, 3);
if( params.speckleWindowSize > 0 )
filterSpeckles(disp, (params.minDisparity - 1)*StereoMatcher::DISP_SCALE, params.speckleWindowSize,
StereoMatcher::DISP_SCALE*params.speckleRange, buffer);
}
static void computeDisparitySGBM( const Mat& img1, const Mat& img2,
Mat& disp1, const StereoSGBMParams& params,
Mat& buffer )
{
const int ALIGN = 16;
const int DISP_SHIFT = StereoMatcher::DISP_SHIFT;
const int DISP_SCALE = (1 << DISP_SHIFT);
const CostType MAX_COST = SHRT_MAX;
int minD = params.minDisparity, maxD = minD + params.numDisparities;
Size SADWindowSize;
SADWindowSize.width = SADWindowSize.height = params.SADWindowSize > 0 ? params.SADWindowSize : 5;
int ftzero = std::max(params.preFilterCap, 15) | 1;
int uniquenessRatio = params.uniquenessRatio >= 0 ? params.uniquenessRatio : 10;
int disp12MaxDiff = params.disp12MaxDiff > 0 ? params.disp12MaxDiff : 1;
int P1 = params.P1 > 0 ? params.P1 : 2, P2 = std::max(params.P2 > 0 ? params.P2 : 5, P1+1);
int k, width = disp1.cols, height = disp1.rows;
int minX1 = std::max(maxD, 0), maxX1 = width + std::min(minD, 0);
int D = maxD - minD, width1 = maxX1 - minX1;
int INVALID_DISP = minD - 1, INVALID_DISP_SCALED = INVALID_DISP*DISP_SCALE;
int SW2 = SADWindowSize.width/2, SH2 = SADWindowSize.height/2;
bool fullDP = params.mode == StereoSGBM::MODE_HH;
int npasses = fullDP ? 2 : 1;
const int TAB_OFS = 256*4, TAB_SIZE = 256 + TAB_OFS*2;
PixType clipTab[TAB_SIZE];
for( k = 0; k < TAB_SIZE; k++ )
clipTab[k] = (PixType)(std::min(std::max(k - TAB_OFS, -ftzero), ftzero) + ftzero);
if( minX1 >= maxX1 )
{
disp1 = Scalar::all(INVALID_DISP_SCALED);
return;
}
CV_Assert( D % 16 == 0 );
// NR - the number of directions. the loop on x below that computes Lr assumes that NR == 8.
// if you change NR, please, modify the loop as well.
int D2 = D+16, NRD2 = NR2*D2;
// the number of L_r(.,.) and min_k L_r(.,.) lines in the buffer:
// for 8-way dynamic programming we need the current row and
// the previous row, i.e. 2 rows in total
const int NLR = 2;
const int LrBorder = NLR - 1;
// for each possible stereo match (img1(x,y) <=> img2(x-d,y))
// we keep pixel difference cost (C) and the summary cost over NR directions (S).
// we also keep all the partial costs for the previous line L_r(x,d) and also min_k L_r(x, k)
size_t costBufSize = width1*D;
size_t CSBufSize = costBufSize*(fullDP ? height : 1);
size_t minLrSize = (width1 + LrBorder*2)*NR2, LrSize = minLrSize*D2;
int hsumBufNRows = SH2*2 + 2;
size_t totalBufSize = (LrSize + minLrSize)*NLR*sizeof(CostType) + // minLr[] and Lr[]
costBufSize*(hsumBufNRows + 1)*sizeof(CostType) + // hsumBuf, pixdiff
CSBufSize*2*sizeof(CostType) + // C, S
width*16*img1.channels()*sizeof(PixType) + // temp buffer for computing per-pixel cost
width*(sizeof(CostType) + sizeof(DispType)) + 1024; // disp2cost + disp2
if( buffer.empty() || !buffer.isContinuous() ||
buffer.cols*buffer.rows*buffer.elemSize() < totalBufSize )
buffer.reserveBuffer(totalBufSize);
// summary cost over different (nDirs) directions
CostType* Cbuf = (CostType*)alignPtr(buffer.ptr(), ALIGN);
CostType* Sbuf = Cbuf + CSBufSize;
CostType* hsumBuf = Sbuf + CSBufSize;
CostType* pixDiff = hsumBuf + costBufSize*hsumBufNRows;
CostType* disp2cost = pixDiff + costBufSize + (LrSize + minLrSize)*NLR;
DispType* disp2ptr = (DispType*)(disp2cost + width);
PixType* tempBuf = (PixType*)(disp2ptr + width);
// add P2 to every C(x,y). it saves a few operations in the inner loops
for(k = 0; k < (int)CSBufSize; k++ )
Cbuf[k] = (CostType)P2;
for( int pass = 1; pass <= npasses; pass++ )
{
int x1, y1, x2, y2, dx, dy;
if( pass == 1 )
{
y1 = 0; y2 = height; dy = 1;
x1 = 0; x2 = width1; dx = 1;
}
else
{
y1 = height-1; y2 = -1; dy = -1;
x1 = width1-1; x2 = -1; dx = -1;
}
CostType *Lr[NLR]={0}, *minLr[NLR]={0};
for( k = 0; k < NLR; k++ )
{
// shift Lr[k] and minLr[k] pointers, because we allocated them with the borders,
// and will occasionally use negative indices with the arrays
// we need to shift Lr[k] pointers by 1, to give the space for d=-1.
// however, then the alignment will be imperfect, i.e. bad for SSE,
// thus we shift the pointers by 8 (8*sizeof(short) == 16 - ideal alignment)
Lr[k] = pixDiff + costBufSize + LrSize*k + NRD2*LrBorder + 8;
memset( Lr[k] - LrBorder*NRD2 - 8, 0, LrSize*sizeof(CostType) );
minLr[k] = pixDiff + costBufSize + LrSize*NLR + minLrSize*k + NR2*LrBorder;
memset( minLr[k] - LrBorder*NR2, 0, minLrSize*sizeof(CostType) );
}
for( int y = y1; y != y2; y += dy )
{
int x, d;
DispType* disp1ptr = disp1.ptr<DispType>(y);
CostType* C = Cbuf + (!fullDP ? 0 : y*costBufSize);
CostType* S = Sbuf + (!fullDP ? 0 : y*costBufSize);
if( pass == 1 ) // compute C on the first pass, and reuse it on the second pass, if any.
{
int dy1 = y == 0 ? 0 : y + SH2, dy2 = y == 0 ? SH2 : dy1;
for( k = dy1; k <= dy2; k++ ) y=0的时候, dy1=0,dy2=2, 3次循环计算3行bt cost, 因为算第一行的window cost需要3行bt cost
{
CostType* hsumAdd = hsumBuf + (std::min(k, height-1) % hsumBufNRows)*costBufSize;
if( k < height )
{
calcPixelCostBT( img1, img2, k, minD, maxD, pixDiff, tempBuf, clipTab, TAB_OFS, ftzero );
memset(hsumAdd, 0, D*sizeof(CostType));
这里x不是实际坐标
minD=12,maxD=92, 实际cost从x=92, 开始，这里x=0是实际坐标x=92
for( x = 0; x <= SW2*D; x += D ) 最左边的点乘3次, 然后再加2点，SW2=2, 是<=,不是<
这样就是第一个完整的hsumAdd了，5个相加，左边一点扩展2点，
{
int scale = x == 0 ? SW2 + 1 : 1;
for( d = 0; d < D; d++ )
hsumAdd[d] = (CostType)(hsumAdd[d] + pixDiff[x + d]*scale);
}
if( y > 0 )
{
const CostType* hsumSub = hsumBuf + (std::max(y - SH2 - 1, 0) % hsumBufNRows)*costBufSize;
const CostType* Cprev = !fullDP || y == 0 ? C : C - costBufSize;
for( x = D; x < width1*D; x += D )
{
注意这里是x=D开始，不是0开始，第一个点的windowed cost只在y=0时算一次，y=1，y=2，...时是沿用y=0时的值，
这个算是BUG还是有意为之？
const CostType* pixAdd = pixDiff + std::min(x + SW2*D, (width1-1)*D);
const CostType* pixSub = pixDiff + std::max(x - (SW2+1)*D, 0);
{
for( d = 0; d < D; d++ )
{
int hv = hsumAdd[x + d] = (CostType)(hsumAdd[x - D + d] + pixAdd[d] - pixSub[d]);
C[x + d] = (CostType)(Cprev[x + d] + hv - hsumSub[x + d]);
}
}
}
}
else
{
for( x = D; x < width1*D; x += D )
{
const CostType* pixAdd = pixDiff + std::min(x + SW2*D, (width1-1)*D);
const CostType* pixSub = pixDiff + std::max(x - (SW2+1)*D, 0);
for( d = 0; d < D; d++ )
hsumAdd[x + d] = (CostType)(hsumAdd[x - D + d] + pixAdd[d] - pixSub[d]);
}
}
}
if( y == 0 )
{
int scale = k == 0 ? SH2 + 1 : 1;
for( x = 0; x < width1*D; x++ )
C[x] = (CostType)(C[x] + hsumAdd[x]*scale); 第0行，乘3倍
}
}
a a a a a|
x x x x x|
x x x x x|
x x x x x|
x x x x x|
---------
b c d e f g
假如滑动窗口5x5
hsumAdd,水平5个相加,hsumAdd放到hsumBuf, 在f位置hsumAdd=b+c+d+e+f,在g位置就是hsumAdd=c+d+e+f+g,
Cprev上一行的Cost，即所有a加所有x，
本次Cost = Cprev - 最上一行的hsumAdd（即5个a相加）+ 最下面一行g位置的hsumAdd(c+d+e+f+g)，
计算本行Cost时要用到上5行的hsumAdd，所以hsumBuf缓存6行的hsumAdd
g位置的hsumAdd(c+d+e+f+g) = f位置的hsumAdd(b+c+d+e+f) + g - b g就是pixAdd，b就是pixSub
//图像高20，窗口大小5
// line 0,0,1 cost 0
// 0,1,2 1
// 1,2,3 2 cost0 *3+cost1+cost2 出4行数据可以求第一个windowed cost0
// 2,3,4 3 cost0*2+cost1+cost2+cost3 出5行数据
// 3,4,5 4 cost0+cost1+cost2+cost3+cost4 出6行数据
// 4,5,6 5
// 5,6,7 6
// 6,7,8 7
// 7,8,9 8
// 8,9,10 9
// 9,10,11 10
// 10,11,12 11
// 11,12,13 12
// 12,13,14 13
// 13,14,15 14
// 14,15,16 15
// 15,16,17 16
// 16,17,18 17
// 17,18,19 18 cost14+cost15+cost16+cost17+cost18 windowed cost16
// 18,19,19 19 cost15+cost16+cost17+cost18+cost19 windowed cost17
// cost15+cost16+cost17+cost18+cost19 windowed cost18
// cost15+cost16+cost17+cost18+cost19 windowed cost19
注意18，19两行的windowed cost并不是cost16+cost17+cost18+cost19*2，cost17+cost18+cost19*3，而是跟17行一样的
// also, clear the S buffer
for( k = 0; k < width1*D; k++ )
S[k] = 0;
}
// clear the left and the right borders
memset( Lr[0] - NRD2*LrBorder - 8, 0, NRD2*LrBorder*sizeof(CostType) );
memset( Lr[0] + width1*NRD2 - 8, 0, NRD2*LrBorder*sizeof(CostType) );
memset( minLr[0] - NR2*LrBorder, 0, NR2*LrBorder*sizeof(CostType) );
memset( minLr[0] + width1*NR2, 0, NR2*LrBorder*sizeof(CostType) );
/*
[formula 13 in the paper]
compute L_r(p, d) = C(p, d) +
min(L_r(p-r, d),
L_r(p-r, d-1) + P1,
L_r(p-r, d+1) + P1,
min_k L_r(p-r, k) + P2) - min_k L_r(p-r, k)
where p = (x,y), r is one of the directions.
we process all the directions at once:
0: r=(-dx, 0)
1: r=(-1, -dy)
2: r=(0, -dy)
3: r=(1, -dy)
4: r=(-2, -dy)
5: r=(-1, -dy*2)
6: r=(1, -dy*2)
7: r=(2, -dy)
*/
for( x = x1; x != x2; x += dx )
{
int xm = x*NR2, xd = xm*D2;
int delta0 = minLr[0][xm - dx*NR2] + P2, delta1 = minLr[1][xm - NR2 + 1] + P2;
int delta2 = minLr[1][xm + 2] + P2, delta3 = minLr[1][xm + NR2 + 3] + P2;
CostType* Lr_p0 = Lr[0] + xd - dx*NRD2;
CostType* Lr_p1 = Lr[1] + xd - NRD2 + D2;
CostType* Lr_p2 = Lr[1] + xd + D2*2;
CostType* Lr_p3 = Lr[1] + xd + NRD2 + D2*3;
Lr_p0[-1] = Lr_p0[D] = Lr_p1[-1] = Lr_p1[D] =
Lr_p2[-1] = Lr_p2[D] = Lr_p3[-1] = Lr_p3[D] = MAX_COST;
CostType* Lr_p = Lr[0] + xd;
const CostType* Cp = C + x*D;
CostType* Sp = S + x*D;
{
int minL0 = MAX_COST, minL1 = MAX_COST, minL2 = MAX_COST, minL3 = MAX_COST;
for( d = 0; d < D; d++ )
{
int Cpd = Cp[d], L0, L1, L2, L3;
L0 = Cpd + std::min((int)Lr_p0[d], std::min(Lr_p0[d-1] + P1, std::min(Lr_p0[d+1] + P1, delta0))) - delta0;
L1 = Cpd + std::min((int)Lr_p1[d], std::min(Lr_p1[d-1] + P1, std::min(Lr_p1[d+1] + P1, delta1))) - delta1;
L2 = Cpd + std::min((int)Lr_p2[d], std::min(Lr_p2[d-1] + P1, std::min(Lr_p2[d+1] + P1, delta2))) - delta2;
L3 = Cpd + std::min((int)Lr_p3[d], std::min(Lr_p3[d-1] + P1, std::min(Lr_p3[d+1] + P1, delta3))) - delta3;
Lr_p[d] = (CostType)L0;
minL0 = std::min(minL0, L0);
Lr_p[d + D2] = (CostType)L1;
minL1 = std::min(minL1, L1);
Lr_p[d + D2*2] = (CostType)L2;
minL2 = std::min(minL2, L2);
Lr_p[d + D2*3] = (CostType)L3;
minL3 = std::min(minL3, L3);
Sp[d] = saturate_cast<CostType>(Sp[d] + L0 + L1 + L2 + L3);
}
minLr[0][xm] = (CostType)minL0;
minLr[0][xm+1] = (CostType)minL1;
minLr[0][xm+2] = (CostType)minL2;
minLr[0][xm+3] = (CostType)minL3;
}
}
if( pass == npasses )
{
for( x = 0; x < width; x++ )
{
disp1ptr[x] = disp2ptr[x] = (DispType)INVALID_DISP_SCALED;
disp2cost[x] = MAX_COST;
}
for( x = width1 - 1; x >= 0; x-- )
{
CostType* Sp = S + x*D;
int minS = MAX_COST, bestDisp = -1;
if( npasses == 1 )
{
int xm = x*NR2, xd = xm*D2;
npasses=1这里进行右左方向的聚合，所以npasses=1共进行5方向聚合，不是4个方向，
npasses=2进行8个方向聚合
int minL0 = MAX_COST;
int delta0 = minLr[0][xm + NR2] + P2;
CostType* Lr_p0 = Lr[0] + xd + NRD2;
Lr_p0[-1] = Lr_p0[D] = MAX_COST;
CostType* Lr_p = Lr[0] + xd;
const CostType* Cp = C + x*D;
{
for( d = 0; d < D; d++ )
{
int L0 = Cp[d] + std::min((int)Lr_p0[d], std::min(Lr_p0[d-1] + P1, std::min(Lr_p0[d+1] + P1, delta0))) - delta0;
Lr_p[d] = (CostType)L0;
minL0 = std::min(minL0, L0);
int Sval = Sp[d] = saturate_cast<CostType>(Sp[d] + L0);
if( Sval < minS )
{
minS = Sval;
bestDisp = d;
}
}
minLr[0][xm] = (CostType)minL0;
}
}
else
{
{
for( d = 0; d < D; d++ )
{
int Sval = Sp[d];
if( Sval < minS )
{
minS = Sval;
bestDisp = d;
}
}
}
}
唯一性，什么叫唯一性，比如bestDisp，最佳视差是50，它对应的Spd=100，其他视差，他们对应的Spd也是100，那就不唯一了，其他视差对应的Spd必须大于小于100一定程度
for( d = 0; d < D; d++ )
{
if( Sp[d]*(100 - uniquenessRatio) < minS*100 && std::abs(bestDisp - d) > 1 )
break;
}
if( d < D )
continue;
d = bestDisp;
int _x2 = x + minX1 - d - minD;
if( disp2cost[_x2] > minS )
{
disp2cost[_x2] = (CostType)minS;
disp2ptr[_x2] = (DispType)(d + minD);
}
if( 0 < d && d < D-1 )
{
// do subpixel quadratic interpolation:
// fit parabola into (x1=d-1, y1=Sp[d-1]), (x2=d, y2=Sp[d]), (x3=d+1, y3=Sp[d+1])
// then find minimum of the parabola.
int denom2 = std::max(Sp[d-1] + Sp[d+1] - 2*Sp[d], 1);
d = d*DISP_SCALE + ((Sp[d-1] - Sp[d+1])*DISP_SCALE + denom2)/(denom2*2);
}
else
d *= DISP_SCALE;
disp1ptr[x + minX1] = (DispType)(d + minD*DISP_SCALE);
}
for( x = minX1; x < maxX1; x++ )
{
左右一致性检测，上面提到的第一篇文章里有描述，
通常意义上左右一致性检测，比如SGM里的，需要以左图为参考计算视差，以右图为参考计算视差，来两遍，再比较一致性，SBGM极大简化了，
假如左图x处的视差为30，那右图x-30处的视差也是30
x=100，视差为30，
disp1ptr[100]=30
disp2ptr[100-30]=30
这样左图每个点遍历下来，并不是每点都对应到disp2ptr[]的一项，而是disp2ptr[]中有些项是没有值的，有些被左图两点遍历到，多次遍历到的就是左右不一致
举个例子,x=90处深度20，x=100处深度30，
disp1ptr[90]=20
disp2ptr[90-20]=20
disp1ptr[100]=30
disp2ptr[100-30]=30
这样disp2ptr[70] 先填30，后又修改成了20，这个点就左右不一致
为啥呢，下面的判断std::abs(disp2ptr[_x] - _d) > disp12MaxDiff成立了，
假如disp12MaxDiff=1
x=100时，d=30, _x=100-d=100-30, disp2ptr[_x]也就是disp2ptr[70]=20, |disp2ptr[70]-d|=|30-20|=10 >disp12MaxDiff ，
// we round the computed disparity both towards -inf and +inf and check
// if either of the corresponding disparities in disp2 is consistent.
// This is to give the computed disparity a chance to look valid if it is.
int d1 = disp1ptr[x];
if( d1 == INVALID_DISP_SCALED )
continue;
int _d = d1 >> DISP_SHIFT;
int d_ = (d1 + DISP_SCALE-1) >> DISP_SHIFT;
int _x = x - _d, x_ = x - d_;
if( 0 <= _x && _x < width && disp2ptr[_x] >= minD && std::abs(disp2ptr[_x] - _d) > disp12MaxDiff &&
0 <= x_ && x_ < width && disp2ptr[x_] >= minD && std::abs(disp2ptr[x_] - d_) > disp12MaxDiff )
disp1ptr[x] = (DispType)INVALID_DISP_SCALED;
}
}
// now shift the cyclic buffers
std::swap( Lr[0], Lr[1] );
std::swap( minLr[0], minLr[1] );
}
}
}
static void calcPixelCostBT( const Mat& img1, const Mat& img2, int y,
int minD, int maxD, CostType* cost,
PixType* buffer, const PixType* tab,
int tabOfs, int , int xrange_min = 0, int xrange_max = DEFAULT_RIGHT_BORDER )
{
int x, c, width = img1.cols, cn = img1.channels();
int minX1 = std::max(maxD, 0), maxX1 = width + std::min(minD, 0);
int D = maxD - minD, width1 = maxX1 - minX1;
//This minX1 & maxX2 correction is defining which part of calculatable line must be calculated
//That is needs of parallel algorithm
xrange_min = (xrange_min < 0) ? 0: xrange_min;
xrange_max = (xrange_max == DEFAULT_RIGHT_BORDER) || (xrange_max > width1) ? width1 : xrange_max;
maxX1 = minX1 + xrange_max;
minX1 += xrange_min;
width1 = maxX1 - minX1;
int minX2 = std::max(minX1 - maxD, 0), maxX2 = std::min(maxX1 - minD, width);
int width2 = maxX2 - minX2;
const PixType *row1 = img1.ptr<PixType>(y), *row2 = img2.ptr<PixType>(y);
网上已有大量文章解释原理,
只添加一点说明，
梯度计算上下扩展一点，
求第一行的梯度，用到第一行，和第二行，
求第二行的梯度，用到第一行，和第二行， 第三行
001
012
123,
...
row1左图，row2右图
PixType *prow1 = buffer + width2*2, *prow2 = prow1 + width*cn*2;
tab += tabOfs;
for( c = 0; c < cn*2; c++ )
{
prow1[width*c] = prow1[width*c + width-1] =
prow2[width*c] = prow2[width*c + width-1] = tab[0]; 梯度，像素，最左和最右的值都是tab[0]， 默认是0x3f
}
int n1 = y > 0 ? -(int)img1.step : 0, s1 = y < img1.rows-1 ? (int)img1.step : 0;
int n2 = y > 0 ? -(int)img2.step : 0, s2 = y < img2.rows-1 ? (int)img2.step : 0;
int minX_cmn = std::min(minX1,minX2)-1;
int maxX_cmn = std::max(maxX1,maxX2)+1;
minX_cmn = std::max(minX_cmn, 1);
maxX_cmn = std::min(maxX_cmn, width - 1);
if( cn == 1 )
{
for( x = minX_cmn; x < maxX_cmn; x++ )
{
n1,n2=-512, s1,s2=512, 前一行和后一行
梯度算子
[ -1,0,1]
[ -2,0,2]
[-1,0,1]
prow1[x] = tab[(row1[x+1] - row1[x-1])*2 + row1[x+n1+1] - row1[x+n1-1] + row1[x+s1+1] - row1[x+s1-1]];
row2也就是右图的，梯度倒过来, 下面原像素也倒过来
prow2[width-1-x] = tab[(row2[x+1] - row2[x-1])*2 + row2[x+n2+1] - row2[x+n2-1] + row2[x+s2+1] - row2[x+s2-1]];
prow1[x+width] = row1[x]; prow1前半部分原像素，后半部分梯度
prow2[width-1-x+width] = row2[x];
prow2[1023] -> 像素0
prow2[512] -> 像素511
prow2[511] -> 梯度0
prow2[0] -> 梯度511
}
}
else
{
for( x = minX_cmn; x < maxX_cmn; x++ )
{
prow1[x] = tab[(row1[x*3+3] - row1[x*3-3])*2 + row1[x*3+n1+3] - row1[x*3+n1-3] + row1[x*3+s1+3] - row1[x*3+s1-3]];
prow1[x+width] = tab[(row1[x*3+4] - row1[x*3-2])*2 + row1[x*3+n1+4] - row1[x*3+n1-2] + row1[x*3+s1+4] - row1[x*3+s1-2]];
prow1[x+width*2] = tab[(row1[x*3+5] - row1[x*3-1])*2 + row1[x*3+n1+5] - row1[x*3+n1-1] + row1[x*3+s1+5] - row1[x*3+s1-1]];
prow2[width-1-x] = tab[(row2[x*3+3] - row2[x*3-3])*2 + row2[x*3+n2+3] - row2[x*3+n2-3] + row2[x*3+s2+3] - row2[x*3+s2-3]];
prow2[width-1-x+width] = tab[(row2[x*3+4] - row2[x*3-2])*2 + row2[x*3+n2+4] - row2[x*3+n2-2] + row2[x*3+s2+4] - row2[x*3+s2-2]];
prow2[width-1-x+width*2] = tab[(row2[x*3+5] - row2[x*3-1])*2 + row2[x*3+n2+5] - row2[x*3+n2-1] + row2[x*3+s2+5] - row2[x*3+s2-1]];
prow1[x+width*3] = row1[x*3];
prow1[x+width*4] = row1[x*3+1];
prow1[x+width*5] = row1[x*3+2];
prow2[width-1-x+width*3] = row2[x*3];
prow2[width-1-x+width*4] = row2[x*3+1];
prow2[width-1-x+width*5] = row2[x*3+2];
}
}
memset( cost + xrange_min*D, 0, width1*D*sizeof(cost[0]) );
buffer -= width-1-maxX2;
cost -= (minX1-xrange_min)*D + minD; // simplify the cost indices inside the loop
1通道要两次-- 第一次梯度，第二次原像素
for( c = 0; c < cn*2; c++, prow1 += width, prow2 += width )
{
int diff_scale = c < cn ? 0 : 2;
// precompute
// v0 = min(row2[x-1/2], row2[x], row2[x+1/2]) and
// v1 = max(row2[x-1/2], row2[x], row2[x+1/2]) and
for( x = width-1-maxX2; x < width-1- minX2; x++ )
{
int v = prow2[x];
int vl = x > 0 ? (v + prow2[x-1])/2 : v; 这个就是半像素的估计值
int vr = x < width-1 ? (v + prow2[x+1])/2 : v;
int v0 = std::min(vl, vr); v0 = std::min(v0, v); 左半，右半，本像素三者最小
int v1 = std::max(vl, vr); v1 = std::max(v1, v); 左半，右半，本像素三者最大
buffer[x] = (PixType)v0;
buffer[x + width2] = (PixType)v1;
}
for( x = minX1; x < maxX1; x++ )
{
int u = prow1[x];
int ul = x > 0 ? (u + prow1[x-1])/2 : u;
int ur = x < width-1 ? (u + prow1[x+1])/2 : u;
int u0 = std::min(ul, ur); u0 = std::min(u0, u);
int u1 = std::max(ul, ur); u1 = std::max(u1, u);
{
for( int d = minD; d < maxD; d++ )
{
int v = prow2[width-x-1 + d];
int v0 = buffer[width-x-1 + d]
int v1 = buffer[width-x-1 + d + width2];
int c0 = std::max(0, u - v1); c0 = std::max(c0, v0 - u);
int c1 = std::max(0, v - u1); c1 = std::max(c1, u0 - v);
}
}
}
}
}
下面是SGBM在fpga上实现，另一篇文章，有兴趣到github下载源代码下来参考一下，
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
