changed mean shift procedure a little

This commit is contained in:
Alexey Spizhevoy
2010-10-11 14:25:30 +00:00
parent 3e840cb798
commit adf7cc20b7
3 changed files with 125 additions and 55 deletions

View File

@@ -163,6 +163,71 @@ namespace imgproc
{
texture<uchar4, 2> tex_meanshift;
__device__ short2 do_mean_shift(int x0, int y0, unsigned char* out,
int out_step, int cols, int rows,
int sp, int sr, int maxIter, float eps)
{
int isr2 = sr*sr;
uchar4 c = tex2D(tex_meanshift, x0, y0 );
// iterate meanshift procedure
for( int iter = 0; iter < maxIter; iter++ )
{
int count = 0;
int s0 = 0, s1 = 0, s2 = 0, sx = 0, sy = 0;
float icount;
//mean shift: process pixels in window (p-sigmaSp)x(p+sigmaSp)
int minx = x0-sp;
int miny = y0-sp;
int maxx = x0+sp;
int maxy = y0+sp;
for( int y = miny; y <= maxy; y++)
{
int rowCount = 0;
for( int x = minx; x <= maxx; x++ )
{
uchar4 t = tex2D( tex_meanshift, x, y );
int norm2 = (t.x - c.x) * (t.x - c.x) + (t.y - c.y) * (t.y - c.y) + (t.z - c.z) * (t.z - c.z);
if( norm2 <= isr2 )
{
s0 += t.x; s1 += t.y; s2 += t.z;
sx += x; rowCount++;
}
}
count += rowCount;
sy += y*rowCount;
}
if( count == 0 )
break;
icount = 1.f/count;
int x1 = __float2int_rz(sx*icount);
int y1 = __float2int_rz(sy*icount);
s0 = __float2int_rz(s0*icount);
s1 = __float2int_rz(s1*icount);
s2 = __float2int_rz(s2*icount);
int norm2 = (s0 - c.x) * (s0 - c.x) + (s1 - c.y) * (s1 - c.y) + (s2 - c.z) * (s2 - c.z);
bool stopFlag = (x0 == x1 && y0 == y1) || (abs(x1-x0) + abs(y1-y0) + norm2 <= eps);
x0 = x1; y0 = y1;
c.x = s0; c.y = s1; c.z = s2;
if( stopFlag )
break;
}
int base = (blockIdx.y * blockDim.y + threadIdx.y) * out_step + (blockIdx.x * blockDim.x + threadIdx.x) * 4 * sizeof(uchar);
*(uchar4*)(out + base) = c;
return make_short2((short)x0, (short)y0);
}
extern "C" __global__ void meanshift_kernel( unsigned char* out, int out_step, int cols, int rows,
int sp, int sr, int maxIter, float eps )
{
@@ -170,63 +235,21 @@ namespace imgproc
int y0 = blockIdx.y * blockDim.y + threadIdx.y;
if( x0 < cols && y0 < rows )
{
int isr2 = sr*sr;
uchar4 c = tex2D(tex_meanshift, x0, y0 );
// iterate meanshift procedure
for( int iter = 0; iter < maxIter; iter++ )
{
int count = 0;
int s0 = 0, s1 = 0, s2 = 0, sx = 0, sy = 0;
float icount;
do_mean_shift(x0, y0, out, out_step, cols, rows, sp, sr, maxIter, eps);
}
//mean shift: process pixels in window (p-sigmaSp)x(p+sigmaSp)
int minx = x0-sp;
int miny = y0-sp;
int maxx = x0+sp;
int maxy = y0+sp;
extern "C" __global__ void meanshiftproc_kernel( unsigned char* outr, int outrstep,
unsigned char* outsp, int outspstep,
int cols, int rows,
int sp, int sr, int maxIter, float eps )
{
int x0 = blockIdx.x * blockDim.x + threadIdx.x;
int y0 = blockIdx.y * blockDim.y + threadIdx.y;
for( int y = miny; y <= maxy; y++)
{
int rowCount = 0;
for( int x = minx; x <= maxx; x++ )
{
uchar4 t = tex2D( tex_meanshift, x, y );
int norm2 = (t.x - c.x) * (t.x - c.x) + (t.y - c.y) * (t.y - c.y) + (t.z - c.z) * (t.z - c.z);
if( norm2 <= isr2 )
{
s0 += t.x; s1 += t.y; s2 += t.z;
sx += x; rowCount++;
}
}
count += rowCount;
sy += y*rowCount;
}
if( count == 0 )
break;
icount = 1.f/count;
int x1 = __float2int_rz(sx*icount);
int y1 = __float2int_rz(sy*icount);
s0 = __float2int_rz(s0*icount);
s1 = __float2int_rz(s1*icount);
s2 = __float2int_rz(s2*icount);
int norm2 = (s0 - c.x) * (s0 - c.x) + (s1 - c.y) * (s1 - c.y) + (s2 - c.z) * (s2 - c.z);
bool stopFlag = (x0 == x1 && y0 == y1) || (abs(x1-x0) + abs(y1-y0) + norm2 <= eps);
x0 = x1; y0 = y1;
c.x = s0; c.y = s1; c.z = s2;
if( stopFlag )
break;
}
int base = (blockIdx.y * blockDim.y + threadIdx.y) * out_step + (blockIdx.x * blockDim.x + threadIdx.x) * 4 * sizeof(uchar);
*(uchar4*)(out + base) = c;
if( x0 < cols && y0 < rows )
{
int basesp = (blockIdx.y * blockDim.y + threadIdx.y) * outspstep + (blockIdx.x * blockDim.x + threadIdx.x) * 2 * sizeof(short);
*(short2*)(outsp + basesp) = do_mean_shift(x0, y0, outr, outrstep, cols, rows, sp, sr, maxIter, eps);
}
}
}
@@ -247,6 +270,20 @@ namespace cv { namespace gpu { namespace improc
cudaSafeCall( cudaThreadSynchronize() );
cudaSafeCall( cudaUnbindTexture( imgproc::tex_meanshift ) );
}
extern "C" void meanShiftProc_gpu(const DevMem2D& src, DevMem2D dstr, DevMem2D dstsp, int sp, int sr, int maxIter, float eps)
{
dim3 grid(1, 1, 1);
dim3 threads(32, 16, 1);
grid.x = divUp(src.cols, threads.x);
grid.y = divUp(src.rows, threads.y);
cudaChannelFormatDesc desc = cudaCreateChannelDesc<uchar4>();
cudaSafeCall( cudaBindTexture2D( 0, imgproc::tex_meanshift, src.ptr, desc, src.cols, src.rows, src.step ) );
imgproc::meanshiftproc_kernel<<< grid, threads >>>( dstr.ptr, dstr.step, dstsp.ptr, dstsp.step, dstr.cols, dstr.rows, sp, sr, maxIter, eps );
cudaSafeCall( cudaThreadSynchronize() );
cudaSafeCall( cudaUnbindTexture( imgproc::tex_meanshift ) );
}
}}}
/////////////////////////////////// drawColorDisp ///////////////////////////////////////////////