mark_vision's profile - activity

This is interesting. The blue channel thing is just because of my copy/paste haste, the original code was correct. Now i try with your solution. What do you mean with "horribly broken", anyway?

Hi there, I'm forced to use the Bumblebee stereo algorithm that use the TriclopsImage and TriclopsImage16 data structures defined by Triclops. I want to convert those structures into a cv::Mat in the fastest way. I've tried with memcpy but it does not work:

Mat srcright(HEIGHT, WIDTH, CV_8UC3);    
for (int i = 0; i < HEIGHT; i++) {
                    memcpy(&srcleft.ptr(i)[0], &rectified_left_color_image.blue[i * rectified_left_color_image.rowinc], WIDTH * sizeof (uchar));
                    memcpy(&srcleft.ptr(i)[1], &rectified_left_color_image.blue[i * rectified_left_color_image.rowinc], WIDTH * sizeof (uchar));
                    memcpy(&srcleft.ptr(i)[2], &rectified_left_color_image.blue[i * rectified_left_color_image.rowinc], WIDTH * sizeof (uchar));
        }

it produces kind of a shrinked image. With the nested for and the point to point assignement with at() is ok but I would prefer a fastest way to do it.

EDIT Now it works, here there's the code:

vector<Mat> channels;

channels.push_back(Mat(HEIGHT, WIDTH, CV_8UC1, rectified_left_color_image.blue));
channels.push_back(Mat(HEIGHT, WIDTH, CV_8UC1, rectified_left_color_image.green));
channels.push_back(Mat(HEIGHT, WIDTH, CV_8UC1, rectified_left_color_image.red));

merge(channels, srcleft);

if I understand what are you asking, try the minMaxLoc function

Hi, I have three Mat of the same size: one is my image and the other two are the (u,v) components of a vector field. I would like to apply a Gaussian filter along each vector (e.g. If the gradient for a pixel is (1,0) the Gaussian kernel is computed only horizontally). How can I do that?

Can I compute something like: p = u*Gx + v*Gy, where Gx is the gaussian blur on the x-axis and Gy the gaussian on the y-axis?

The idea is that I have some contours and I want to compute the gaussian blur on the normal of the contour instead of blindly using a circular or elliptical surrounding of pixels for the computation.