I am trying to migrate some code from Matlab to Opencv and need an exact replica of the gradient function. I have tried the cv::Sobel function but for some reason the values in the resulting cv::Mat are not the same as the values in the Matlab version. I need the X and Y gradient in separate matrices for further calculations.
Any workaround that could achieve this would be great
Sobel can only compute the second derivative of the image pixel which is not what we want.
(f(i+1,j) + f(i-1,j) - 2f(i,j)) / 2
What we want is
(f(i+i,j)-f(i-1,j)) / 2
So we need to apply
Mat kernelx = (Mat_<float>(1,3)<<-0.5, 0, 0.5);
Mat kernely = (Mat_<float>(3,1)<<-0.5, 0, 0.5);
filter2D(src, fx, -1, kernelx)
filter2D(src, fy, -1, kernely);
Matlab treats border pixels differently from inner pixels. So the code above is wrong at the border values. One can use BORDER_CONSTANT to extent the border value out with a constant number, unfortunately the constant number is -1 by OpenCV and can not be changed to 0 (which is what we want).
So as to border values, I do not have a very neat answer to it. Just try to compute the first derivative by hand...
You have to call Sobel 2 times, with arguments:
xorder = 1, yorder = 0
and
xorder = 0, yorder = 1
You have to select the appropriate kernel size.
See documentation
It might still be that the MatLab implementation was different, ideally you should retrieve which kernel was used there...
Edit:
If you need to specify your own kernel, you can use the more generic filter2D. Your destination depth will be CV_16S (16bit signed).
Matlab computes the gradient differently for interior rows and border rows (the same is true for the columns of course). At the borders, it is a simple forward difference gradY(1) = row(2) - row(1). The gradient for interior rows is computed by the central difference gradY(2) = (row(3) - row(1)) / 2.
I think you cannot achieve the same result with just running a single convolution filter over the whole matrix in OpenCV. Use cv::Sobel() with ksize = 1, then treat the borders (either manually or by applying a [ 1 -1 ] filter).
Pei's answer is partly correct. Matlab uses these calculations for the borders:
G(:,1) = A(:,2) - A(:,1);
G(:,N) = A(:,N) - A(:,N-1);
so used the following opencv code to complete the gradient:
static cv::Mat kernelx = (cv::Mat_<double>(1, 3) << -0.5, 0, 0.5);
static cv::Mat kernely = (cv::Mat_<double>(3, 1) << -0.5, 0, 0.5);
cv::Mat fx, fy;
cv::filter2D(Image, fx, -1, kernelx, cv::Point(-1, -1), 0, cv::BORDER_REPLICATE);
cv::filter2D(Image, fy, -1, kernely, cv::Point(-1, -1), 0, cv::BORDER_REPLICATE);
fx.col(fx.cols - 1) *= 2;
fx.col(0) *= 2;
fy.row(fy.rows - 1) *= 2;
fy.row(0) *= 2;
Jorrit's answer is partly correct.
In some cases, the value of the directional derivative may be negative, and MATLAB will retain these negative numbers, but OpenCV Mat will set the negative number to 0.
Related
My binary image has rectangular rotated objects of known size on it. I'd like to get the object inclination using axis-aligned bounding box that MATLAB's regionprops returns. What are my suggestions:
Let bounding box width be W, side of rectangle be C and inclination alpha
Then
Using Weierstrass substitution
After some simplification:
Solving the equation for tan(alpha/2) with
For any nonzero inclination discriminant is positive.
Logic seems to be OK, so as math. Could you please point where I make a mistake, or what is a better way to get inclination?
Here is corresponding MATLAB code:
img = false(25,25);
img(5:16,5:16) = true;
rot_img = imrotate(img, 30, 'crop');
props = regionprops(bwlabel(rot_img),'BoundingBox');
bbox = cat(1,props.BoundingBox);
w = bbox(3);
h = 12;
a = -1*(1+w/h); b = 2; c = 1 - w/h;
D = b^2 - 4*a*c;
alpha = 2*atand((-b + sqrt(D))/(2*a));
%alpha = 25.5288
EDIT Thank you for trigonometry hints. They significantly simplify the calculations, but they give wrong answer. As I now understand, the question is asked in wrong way. The thing I really need is finding inclination of short lines (10-50 pixels) with high accuracy (+/- 0.5 deg), the lines' position is out of interest.
The approach used in the question and answers show better accuracy for long lines, for c = 100 error is less than 0.1 degree. That means we're into rasterization error here, and need subpixel accuracy. At the moment I have only one algorithm that solves the problem - Radon transform, but I hope you can recommend something else.
p = bwperim(rot_img);
theta=0:0.1:179.9;
[R,xp] = radon(p,theta); %Radon transform of contours
a=imregionalmax(R,true(3,3)); %Regional maxima of the transform
[r,c]=find(a); idx=sub2ind(size(a),r,c); maxvals=R(idx);
[val,midx]=sort(maxvals,'descend'); %Choose 4 highest maxima
mean(rem(theta(c(midx(1:4))),90)) %And average corresponding angles
%29.85
If rectangle is square:
w/c=sin(a)+cos(a)
(w/c)^2=1+sin(2a)
sin(2a)=(w/c)^2-1
a=0.5*arcsin((w/c)^2-1)
May be use regionprops function with 'Orientation' option...
This question already has answers here:
Matlab: separate connected components
(2 answers)
Closed 5 years ago.
I want to understand how imfindcircles works, so I created a simple image with black background and a single white circle. The image is 640x480 and the circle has a diameter of 122 pixels:
I tried to use imfindcircles to detect the circle, I have tried various modes of the image, as uint8 RGB, uint8 grayscale, double grayscale and the reversed image, in all those forms, and with various values for minR and maxR. I got no result in all cases:
minR = 40;
maxR = 80;
Irgb = imread('example_circle.png');
Irgbr = 255 - Irgb;
I = rgb2gray(Irgb);
Ir = 255 - I;
Id = double(I)/255;
Ird = 1 - Id;
[c1,r1] = imfindcircles(I,[minR maxR]);
[c2,r2] = imfindcircles(Ir,[minR maxR]);
[c3,r3] = imfindcircles(Id,[minR maxR]);
[c4,r4] = imfindcircles(Ird,[minR maxR]);
[c5,r5] = imfindcircles(Irgb,[minR maxR]);
[c6,r6] = imfindcircles(Irgbr,[minR maxR]);
disp([length(r1) length(r2) length(r3) length(r4) length(r5) length(r6)]);
The output is:
0 0 0 0 0 0
How am I supposed to use the function to find the circle?
The imfindcircles function has a 'Sensitivity' parameter:
As you increase the sensitivity factor, imfindcircles detects more circular objects, including weak and partially obscured circles. Higher sensitivity values also increase the risk of false detection.
By setting the Sensitivity to a higher value, you get more potential circles. You could tune this parameter to always give you one circle, e.g. 0.95 seems to work fine in this specific case. This is probably not very robust though.
[c1, r1] = imfindcircles(Irgb,[40,80], 'Sensitivity', 0.95)
If you know that there will always be exactly one circle, you can set the Sensitivity to 1, which returns all potential circles. Then, use the metric output, which gives you the calculated strength of a detected circle. As you know there will be exactly one circle, you can just take the strongest one, which is always the first row.
[c, r] = imfindcircles(Irgb,[40,80], 'Sensitivity', 1);
c1 = c(1,:);
r1 = r(1,:);
I want to apply Temporal Median Filter to a depth map video to ensure temporal consistency and prevent the flickering effect.
Thus, I am trying to apply the filter on all video frames at once by:
First loading all frames,
%%% Read video sequence
numfrm = 5;
infile_name = 'depth_map_1920x1088_80fps.yuv';
width = 1920; %xdim
height = 1088; %ydim
fid_in = fopen(infile_name, 'rb');
[Yd, Ud, Vd] = yuv_import(infile_name,[width, height],numfrm);
fclose(fid_in);
then creating a 3-D depth matrix (height x width x number-of-frames),
%%% Build a stack of images from the video sequence
stack = zeros(height, width, numfrm);
for i=1:numfrm
RGB = yuv2rgb(Yd{i}, Ud{i}, Vd{i});
RGB = RGB(:, :, 1);
stack(:,:,i) = RGB;
end
and finally applying the 1-D median filter along the third direction (time)
temp = medfilt1(stack);
However, for some reason this is not working. When I try to view each frame, I get white images.
frame1 = temp(:,:,1);
imshow(frame1);
Any help would be appreciated!
My guess is that this is actually working but frame1 is of class double and contains values, e.g. between 0 and 255. As imshow represents double images by default on a [0,1] scale, you obtain a white, saturated image.
I would therefore suggest:
caxis auto
after imshow to fix the display problem.
Best,
I am working on an application where I need to rectify an image taken from a mobile camera platform. The platform measures roll, pitch and yaw angles, and I want to make it look like the image is taken from directly above, by some sort of transform from this information.
In other words, I want a perfect square lying flat on the ground, photographed from afar with some camera orientation, to be transformed, so that the square is perfectly symmetrical afterwards.
I have been trying to do this through OpenCV(C++) and Matlab, but I seem to be missing something fundamental about how this is done.
In Matlab, I have tried the following:
%% Transform perspective
img = imread('my_favourite_image.jpg');
R = R_z(yaw_angle)*R_y(pitch_angle)*R_x(roll_angle);
tform = projective2d(R);
outputImage = imwarp(img,tform);
figure(1), imshow(outputImage);
Where R_z/y/x are the standard rotational matrices (implemented with degrees).
For some yaw-rotation, it all works just fine:
R = R_z(10)*R_y(0)*R_x(0);
Which gives the result:
If I try to rotate the image by the same amount about the X- or Y- axes, I get results like this:
R = R_z(10)*R_y(0)*R_x(10);
However, if I rotate by 10 degrees, divided by some huge number, it starts to look OK. But then again, this is a result that has no research value what so ever:
R = R_z(10)*R_y(0)*R_x(10/1000);
Can someone please help me understand why rotating about the X- or Y-axes makes the transformation go wild? Is there any way of solving this without dividing by some random number and other magic tricks? Is this maybe something that can be solved using Euler parameters of some sort? Any help will be highly appreciated!
Update: Full setup and measurements
For completeness, the full test code and initial image has been added, as well as the platforms Euler angles:
Code:
%% Transform perspective
function [] = main()
img = imread('some_image.jpg');
R = R_z(0)*R_y(0)*R_x(10);
tform = projective2d(R);
outputImage = imwarp(img,tform);
figure(1), imshow(outputImage);
end
%% Matrix for Yaw-rotation about the Z-axis
function [R] = R_z(psi)
R = [cosd(psi) -sind(psi) 0;
sind(psi) cosd(psi) 0;
0 0 1];
end
%% Matrix for Pitch-rotation about the Y-axis
function [R] = R_y(theta)
R = [cosd(theta) 0 sind(theta);
0 1 0 ;
-sind(theta) 0 cosd(theta) ];
end
%% Matrix for Roll-rotation about the X-axis
function [R] = R_x(phi)
R = [1 0 0;
0 cosd(phi) -sind(phi);
0 sind(phi) cosd(phi)];
end
The initial image:
Camera platform measurements in the BODY coordinate frame:
Roll: -10
Pitch: -30
Yaw: 166 (angular deviation from north)
From what I understand the Yaw-angle is not directly relevant to the transformation. I might, however, be wrong about this.
Additional info:
I would like specify that the environment in which the setup will be used contains no lines (oceanic photo) that can reliably used as a reference (the horizon will usually not be in the picture). Also the square in the initial image is merely used as a measure to see if the transformation is correct, and will not be there in a real scenario.
So, this is what I ended up doing: I figured that unless you are actually dealing with 3D images, rectifying the perspective of a photo is a 2D operation. With this in mind, I replaced the z-axis values of the transformation matrix with zeros and ones, and applied a 2D Affine transformation to the image.
Rotation of the initial image (see initial post) with measured Roll = -10 and Pitch = -30 was done in the following manner:
R_rotation = R_y(-60)*R_x(10);
R_2d = [ R_rot(1,1) R_rot(1,2) 0;
R_rot(2,1) R_rot(2,2) 0;
0 0 1 ]
This implies a rotation of the camera platform to a virtual camera orientation where the camera is placed above the scene, pointing straight downwards. Note the values used for roll and pitch in the matrix above.
Additionally, if rotating the image so that is aligned with the platform heading, a rotation about the z-axis might be added, giving:
R_rotation = R_y(-60)*R_x(10)*R_z(some_heading);
R_2d = [ R_rot(1,1) R_rot(1,2) 0;
R_rot(2,1) R_rot(2,2) 0;
0 0 1 ]
Note that this does not change the actual image - it only rotates it.
As a result, the initial image rotated about the Y- and X-axes looks like:
The full code for doing this transformation, as displayed above, was:
% Load image
img = imread('initial_image.jpg');
% Full rotation matrix. Z-axis included, but not used.
R_rot = R_y(-60)*R_x(10)*R_z(0);
% Strip the values related to the Z-axis from R_rot
R_2d = [ R_rot(1,1) R_rot(1,2) 0;
R_rot(2,1) R_rot(2,2) 0;
0 0 1 ];
% Generate transformation matrix, and warp (matlab syntax)
tform = affine2d(R_2d);
outputImage = imwarp(img,tform);
% Display image
figure(1), imshow(outputImage);
%*** Rotation Matrix Functions ***%
%% Matrix for Yaw-rotation about the Z-axis
function [R] = R_z(psi)
R = [cosd(psi) -sind(psi) 0;
sind(psi) cosd(psi) 0;
0 0 1];
end
%% Matrix for Pitch-rotation about the Y-axis
function [R] = R_y(theta)
R = [cosd(theta) 0 sind(theta);
0 1 0 ;
-sind(theta) 0 cosd(theta) ];
end
%% Matrix for Roll-rotation about the X-axis
function [R] = R_x(phi)
R = [1 0 0;
0 cosd(phi) -sind(phi);
0 sind(phi) cosd(phi)];
end
Thank you for the support, I hope this helps someone!
I think you can derive transformation this way:
1) Let you have four 3d-points A(-1,-1,0), B(1,-1,0), C(1,1,0) and D(-1,1,0). You can take any 4 noncollinear points. They not related to image.
2) You have transformation matrix, so you can set your camera by multiplying points coords by transformation matrix. And you'll get 3d coords relative to camera position/direction.
3) You need to get projection of your points to screen plane. The simpliest way is to use ortographic projection (simply ignore depth coordinate). On this stage you've got 2D projections of transformed points.
4) Once you have 2 sets of 2D points coordinates (the set from step 1 without 3-rd coordinate and the set from step 3), you can compute homography matrix in standard way.
5) Apply inverse homograhy transformation to your image.
You need to estimate a homography. For an off-the-shelf Matlab solution, see function vgg_H_from_x_lin.m from http://www.robots.ox.ac.uk/~vgg/hzbook/code/ .
For the theory dig into a Computer Vision textbook, such as the one available freely at http://szeliski.org/Book/ or in Chapter 3 of http://programmingcomputervision.com/downloads/ProgrammingComputerVision_CCdraft.pdf
Maybe my answer is not correct due to my mis-understanding of the camera parameters, but I was wondering whether the Yaw/Pitch/Roll is relative to the position of your object. I used the formula of general rotations, and my code is below (the rotation functions R_x, R_y, and R_z were copied from yours, I didn't paste them here)
close all
file='http://i.stack.imgur.com/m5e01.jpg'; % original image
I=imread(file);
R_rot = R_x(-10)*R_y(-30)*R_z(166);
R_rot = inv(R_rot);
R_2d = [ R_rot(1,1) R_rot(1,2) 0;
R_rot(2,1) R_rot(2,2) 0;
0 0 1 ];
T = maketform('affine',R_2d);
transformedI = imtransform(I,T);
figure, imshow(I), figure, imshow(transformedI)
The result:
This indicates that you still need some rotation operation to get the 'correct' alignment in your mind (but probably not necessary the correct position in the camera's mind).
So I change R_rot = inv(R_rot); to R_rot = inv(R_rot)*R_x(-5)*R_y(25)*R_z(180);, and now it gave me:
Looks better like what you want.
Thanks.
I'm working on an image processing project. I have a grayscale image and detected edges with Canny edge detection. Now I would like to manipulate the result by filtering the unnecessary edges. I would like to keep the edges which are close to horizontal and delete edges which are close to vertical.
How can I delete the close to vertical edges?
One option is to use half of a Sobel operator. The full algorithm finds horizontal and vertical edges, then combines them. You are only interested in horizontal edges, so just compute that part (which is Gy in the Wikipedia article).
You may also want to threshold the result to get a black and white image instead of shades of gray.
You could apply this technique to the original grayscale image or the result of the Canny edge detection.
It depends on how cost-intensive it is allowed to be. One easy way to do would be:
(1) Convolute your image with Sobel-Filters (gives Dx, Dy).
For each canny-edge-pixel:
(2) Normalize (Dx, Dy), s.t. in every pixel you have the direction of your edge.
(3) Compute the inner products with the direction you want to remove (in your case (0,1)).
(4) If the absolut value of the inner product is smaller than some threshold remove the pixel.
Example:
img = ...;
canny_img = ...;
removeDir = [0;1];
% convolute with sobel masks
sobelX = [1, 0, -1; 2, 0, -2; 1, 0, -1];
sobelY = sobelX';
DxImg = conv2(img,sobelX,'same');
DyImg = conv2(img,sobelY,'same');
% for each canny-edge-pixel:
for lin = 1:size(img,1) % <-> y
for col = 1:size(img,2) % <-> x
if canny_img(lin,col)
% normalize direction
normDir = [DxImg(lin,col); DyImg(lin,col)];
normDir = normDir / norm(normDir,2);
% inner product
innerP = normDir' * removeDir;
% remove edge?
if abs(innerP) < cos(45/180*pi) % 45° threshold
canny_img(lin,col) = 0;
end
end
end
end
You can optimize it a lot due to your requirements.