void Convolve(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage kernel)
{
// Determine the shape of the kernel and source image
CShape kShape = kernel.Shape();
CShape sShape = src.Shape();
// Allocate the result, if necessary
dst.ReAllocate(sShape, false);
if (sShape.width * sShape.height * sShape.nBands == 0)
return;
// Do the convolution
for (int y = 0; y < sShape.height; y++)
for (int x = 0; x < sShape.width; x++)
for (int c = 0; c < sShape.nBands; c++)
{
double sum = 0;
for (int k = 0; k < kShape.height; k++)
for (int l = 0; l < kShape.width; l++)
if ((x-kernel.origin[0]+k >= 0) && (x-kernel.origin[0]+k < sShape.width) && (y-kernel.origin[1]+l >= 0) && (y-kernel.origin[1]+l < sShape.height))
sum += kernel.Pixel(k,l,0) * src.Pixel(x-kernel.origin[0]+k,y-kernel.origin[1]+l,c);
dst.Pixel(x,y,c) = (T) __max(dst.MinVal(), __min(dst.MaxVal(), sum));
}
}
void WarpLocal(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage uv, bool relativeCoords,
EWarpInterpolationMode interp, float cubicA)
{
// Check that dst is of the right shape
CShape sh(uv.Shape().width, uv.Shape().height, src.Shape().nBands);
dst.ReAllocate(sh);
// Allocate a row buffer for coordinates
int n = sh.width;
std::vector<float> rowBuf;
rowBuf.resize(n*2);
// Precompute the cubic interpolant
if (interp == eWarpInterpCubic)
InitializeCubicLUT(cubicA);
// Process each row
for (int y = 0; y < sh.height; y++)
{
float *uvP = &uv .Pixel(0, y, 0);
float *xyP = (relativeCoords) ? &rowBuf[0] : uvP;
T *dstP = &dst.Pixel(0, y, 0);
// Convert to absolute coordinates if necessary
if (relativeCoords)
{
for (int x = 0; x < n; x++)
{
xyP[2*x+0] = x + uvP[2*x+0];
xyP[2*x+1] = y + uvP[2*x+1];
}
}
// Resample the line
WarpLine(src, dstP, xyP, n, sh.nBands, interp, src.MinVal(), src.MaxVal());
}
}
void Convolve(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage kernel)
{
// Allocate the result, if necessary
dst.ReAllocate(src.Shape(), false);
// Create the IPL images
IplImage* srcImage = IPLCreateImage(src);
IplImage* dstImage = IPLCreateImage(dst);
// Call the convolution code
if (typeid(T) == typeid(float)) {
IplConvKernelFP* iplKernel = IPLConvKernelFP(kernel);
iplConvolve2DFP(srcImage, dstImage, &iplKernel, 1, IPL_SUM);
iplDeleteConvKernelFP(iplKernel);
} else {
IplConvKernel* iplKernel = IPLConvKernel(kernel);
iplConvolve2D(srcImage, dstImage, &iplKernel, 1, IPL_SUM);
iplDeleteConvKernel(iplKernel);
}
iplDeallocate(srcImage, IPL_IMAGE_HEADER);
iplDeallocate(dstImage, IPL_IMAGE_HEADER);
}
void testGetFloatFromMatrix()
{
CImageOf<double> img = GetImageFromMatrix((double *)sobelX, 3, 3);
bool areSame = true;
for (int i = 0; i < 3; i++)
{
for (int j = 0; j < 3; j++)
{
float imgPixel = (float)img.Pixel(j, i, 0);
float sobelPixel = (float)sobelX[i * 3 + j];
if (img.Pixel(j, i, 0) != sobelX[i * 3 + j])
{
areSame = false;
}
}
}
if (areSame)
printf("Same\n");
else
printf("Not the same\n");
}
void ConvolveSeparable(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage x_kernel, CFloatImage y_kernel,
int subsample)
{
// Allocate the result, if necessary
CShape dShape = src.Shape();
if (subsample > 1) {
dShape.width = (dShape.width + subsample - 1) / subsample;
dShape.height = (dShape.height + subsample - 1) / subsample;
}
dst.ReAllocate(dShape, false);
// Allocate the intermediate (or final) result
CImageOf<T> tmpImg;
if (subsample > 1)
tmpImg.ReAllocate(src.Shape());
CImageOf<T>& tmp = (subsample > 1) ? tmpImg : dst;
// Create the IPL images
IplImage* srcImage = IPLCreateImage(src);
IplImage* dstImage = IPLCreateImage(tmp);
srcImage->alphaChannel = 0; // convolve the A channel also
dstImage->alphaChannel = 0; // convolve the A channel also
// Call the convolution code
if (typeid(T) == typeid(float)) {
IplConvKernelFP* xKernel = IPLConvKernelFP(x_kernel, false);
IplConvKernelFP* yKernel = IPLConvKernelFP(y_kernel, true);
iplConvolveSep2DFP(srcImage, dstImage, xKernel, yKernel);
iplDeleteConvKernelFP(xKernel);
iplDeleteConvKernelFP(yKernel);
} else {
IplConvKernel* xKernel = IPLConvKernel(x_kernel, false);
IplConvKernel* yKernel = IPLConvKernel(y_kernel, true);
iplConvolveSep2D(srcImage, dstImage, xKernel, yKernel);
iplDeleteConvKernel(xKernel);
iplDeleteConvKernel(yKernel);
}
iplDeallocate(srcImage, IPL_IMAGE_HEADER);
iplDeallocate(dstImage, IPL_IMAGE_HEADER);
// Downsample if necessary
if (subsample > 1) {
for (int y = 0; y < dShape.height; y++) {
T* sPtr = &tmp.Pixel(0, y * subsample, 0);
T* dPtr = &dst.Pixel(0, y, 0);
int nB = dShape.nBands;
for (int x = 0; x < dShape.width; x++) {
for (int b = 0; b < nB; b++)
dPtr[b] = sPtr[b];
sPtr += subsample * nB;
dPtr += nB;
}
}
}
}
static void FillRowBuffer(T buf[], CImageOf<T>& src, CFloatImage& kernel, int k, int n)
{
// Compute the real row address
CShape sShape = src.Shape();
int nB = sShape.nBands;
int k0 = TrimIndex(k + kernel.origin[1], src.borderMode, sShape.height);
if (k0 < 0)
{
memset(buf, 0, n * sizeof(T));
return;
}
// Fill the row
T* srcP = &src.Pixel(0, k0, 0);
int m = n / nB;
for (int l = 0; l < m; l++, buf += nB)
{
int l0 = TrimIndex(l + kernel.origin[0], src.borderMode, sShape.width);
if (l0 < 0)
memset(buf, 0, nB * sizeof(T));
else
memcpy(buf, &srcP[l0*nB], nB * sizeof(T));
}
}
void ConvolveSeparable(CImageOf<T> src, CImageOf<T>& dst,
CFloatImage x_kernel, CFloatImage y_kernel,
float scale, float offset,
int decimate, int interpolate)
{
// Allocate the result, if necessary
CShape dShape = src.Shape();
if (decimate > 1)
{
dShape.width = (dShape.width + decimate-1) / decimate;
dShape.height = (dShape.height + decimate-1) / decimate;
}
dst.ReAllocate(dShape, false);
// Allocate the intermediate images
CImageOf<T> tmpImg1(src.Shape());
//CImageOf<T> tmpImgCUDA(src.Shape());
CImageOf<T> tmpImg2(src.Shape());
// Create a proper vertical convolution kernel
CFloatImage v_kernel(1, y_kernel.Shape().width, 1);
for (int k = 0; k < y_kernel.Shape().width; k++)
v_kernel.Pixel(0, k, 0) = y_kernel.Pixel(k, 0, 0);
v_kernel.origin[1] = y_kernel.origin[0];
#ifdef RUN_ON_GPU
// Modifications for integrating CUDA kernels
BinomialFilterType type;
profilingTimer->startTimer();
// CUDA Convolve
switch (x_kernel.Shape().width)
{
case 3:
type = BINOMIAL6126;
break;
case 5:
type = BINOMIAL14641;
break;
default:
// Unsupported kernel case
throw CError("Convolution kernel Unknown");
assert(false);
}
// Skip copy if decimation is not required
if (decimate != 1) CudaConvolveXY(src, tmpImg2, type);
else CudaConvolveXY(src, dst, type);
printf("\nGPU convolution time = %f ms\n", profilingTimer->stopAndGetTimerValue());
#else
profilingTimer->startTimer();
//VerifyComputedData(&tmpImg2.Pixel(0, 0, 0), &tmpImgCUDA.Pixel(0, 0, 0), 7003904);
// Perform the two convolutions
Convolve(src, tmpImg1, x_kernel, 1.0f, 0.0f);
Convolve(tmpImg1, tmpImg2, v_kernel, scale, offset);
printf("\nCPU Convolution time = %f ms\n", profilingTimer->stopAndGetTimerValue());
#endif
profilingTimer->startTimer();
// Downsample or copy
// Skip decimate and recopy if not required
#ifdef RUN_ON_GPU
if (decimate != 1)
{
#endif
for (int y = 0; y < dShape.height; y++)
{
T* sPtr = &tmpImg2.Pixel(0, y * decimate, 0);
T* dPtr = &dst.Pixel(0, y, 0);
int nB = dShape.nBands;
for (int x = 0; x < dShape.width; x++)
{
for (int b = 0; b < nB; b++)
dPtr[b] = sPtr[b];
sPtr += decimate * nB;
dPtr += nB;
}
}
#ifdef RUN_ON_GPU
}
#endif
printf("\nDecimate/Recopy took = %f ms\n", profilingTimer->stopAndGetTimerValue());
}
extern void ForwardWarp(CImageOf<T>& src, CImageOf<T>& dst, CFloatImage& disp,
float d_scale, bool line_interpolate, float disp_gap)
{
// Warps src into dst using disparities disp.
// Each disparity is scaled by d_scale
// Note that "empty" pixels are left at their original value
CShape sh = src.Shape();
int w = sh.width, h = sh.height, n_bands = sh.nBands;
float round_offset = (typeid(T) == typeid(float)) ? 0.0f : 0.5f;
if (! sh.SameIgnoringNBands(disp.Shape()))
throw CError("ForwardWarp: disparity image has wrong size");
if (sh != dst.Shape())
dst.ReAllocate(sh);
// Optional clipping (if necessary)
CFloatImage flt;
T minVal = dst.MinVal();
T maxVal = dst.MaxVal();
if (minVal <= flt.MinVal() && maxVal >= flt.MaxVal())
minVal = maxVal = 0;
for (int y = 0; y < h; y++)
{
// determine correct warping direction
int xstart = (d_scale>0 ? 0 : w-1);
int xend = (d_scale>0 ? w : -1 );
int xincr = (d_scale>0 ? 1 : -1 );
float *dp = &disp.Pixel(0, y, 0);
T *ps = &src .Pixel(0, y, 0);
T *pd = &dst .Pixel(0, y, 0);
for (int x = xstart; x != xend; x += xincr)
{
// determine if a line should be drawn
int x2 = x + xincr;
float d_diff = fabs(dp[x] - dp[x2]);
bool draw_line = line_interpolate && (x2 != xend) &&
(d_diff < disp_gap);
// scaled disparity:
float d = d_scale * dp[x];
// line drawing
if (draw_line)
{
float d2 = d_scale * dp[x2];
if (xincr > 0)
draw_intensity_line(&ps[x * n_bands], &ps[x2 * n_bands], pd,
x - d, x2 - d2, w, n_bands, round_offset,
minVal, maxVal);
else
draw_intensity_line(&ps[x2 * n_bands], &ps[x * n_bands], pd,
x2 - d, x - d2, w, n_bands, round_offset,
minVal, maxVal);
continue;
}
// splatting
int xx = x - ROUND(d);
if (xx >= 0 && xx < w)
memcpy(&pd[xx * n_bands], &ps[x * n_bands],
n_bands*sizeof(T));
}
}
}
extern void InverseWarp(CImageOf<T>& src, CImageOf<T>& dst, CFloatImage& disp,
float d_scale, float disp_gap, int order)
{
// Warps src into dst using disparities disp.
// Each disparity is scaled by d_scale
// Note that "empty" pixels are left at their original value
CShape sh = src.Shape();
int w = sh.width, h = sh.height, n_bands = sh.nBands;
int n = w * n_bands;
if (! sh.SameIgnoringNBands(disp.Shape()))
throw CError("InverseWarp: disparity image has wrong size");
if (sh != dst.Shape())
dst.ReAllocate(sh);
// Optional forward warped depth map if checking for visibility
CFloatImage fwd, fwd_tmp;
if (disp_gap > 0.0f)
{
ScaleAndOffset(disp, fwd_tmp, d_scale, 0.0f);
fwd.ReAllocate(disp.Shape());
fwd.FillPixels(-9999.0f);
ForwardWarp(fwd_tmp, fwd, disp, d_scale, true, disp_gap);
}
// Allocate line buffers
std::vector<float> src_buf, dst_buf, dsp_buf;
src_buf.resize(n);
dst_buf.resize(n);
dsp_buf.resize(n);
CFloatImage fimg; // dummy, used for MinVal(), MaxVal()
for (int y = 0; y < h; y++)
{
// Set up (copy) the line buffers
ScaleAndOffsetLine(&src .Pixel(0, y, 0), &src_buf[0], n,
1.0f, 0.0f, fimg.MinVal(), fimg.MaxVal());
ScaleAndOffsetLine(&disp.Pixel(0, y, 0), &dsp_buf[0], w,
d_scale, 0.0f, 0.0f, 0.0f);
ScaleAndOffsetLine(&dst .Pixel(0, y, 0), &dst_buf[0], n,
1.0f, 0.0f, fimg.MinVal(), fimg.MaxVal());
// Forward warp the depth map
float *fwd_buf = (disp_gap > 0.0f) ? &fwd.Pixel(0, y, 0) : 0;
// Process (warp) the line
InverseWarpLine(&src_buf[0], &dst_buf[0], &dsp_buf[0],
w, n_bands, order, fwd_buf, disp_gap);
// Convert back to native type
T minVal = dst.MinVal();
T maxVal = dst.MaxVal();
float offset = (typeid(T) == typeid(float)) ? 0.0f : 0.5; // rounding
if (minVal <= fimg.MinVal() && maxVal >= fimg.MaxVal())
minVal = maxVal = 0;
ScaleAndOffsetLine(&dst_buf[0], &dst.Pixel(0, y, 0), n,
1.0f, offset, minVal, maxVal);
}
}
