I am trying to implement the snake algorithm for active contour using C++ and OpenCV 3. I am working with the version that uses the gradient descent. As base test I am trying to draw a contour of a lip. This is the base image.
This is the evolution of the contour without external forces (alpha = 0.001, beta = 3, step-size=0.3).
When I add the external force, this is the result.
As external force I have used just the edge detection with Sobel derivative.
This is the code I use for points update.
array<Mat, 2> edges = edgeMatrices(croppedImage);
const float ALPHA = 0.001, BETA = 3, GAMMA = 0.3, // Gamma is step size.
a = GAMMA * ALPHA, b = GAMMA * BETA;
const uint16_t CYCLES = 1000;
const float p = b, q = -a - 4 * b, r = 1 + 2 * a + 6 * b;
Mat pMatrix = pentadiagonalMatrix(POINTS_NUM, p, q, r).inv();
for (uint16_t i = 0; i < CYCLES; ++i) {
// Extract the x and y derivatives for current points.
auto externalForces = external(edges, x, y);
x = pMatrix * (x + GAMMA * externalForces[0]);
y = pMatrix * (y + GAMMA * externalForces[1]);
// Draw the points.
if (i % 200 == 0 && i > 0)
drawPoints(croppedImage, x, y, { 0.2f * i, 0.2f * i, 0 });
}
This is the code for computing the derivatives.
array<Mat, 2> edgeMatrices(Mat &img) {
// Convert image.
Mat gray;
cvtColor(img, gray, COLOR_BGR2GRAY);
// Apply scharr filter.
Mat grad_x, grad_y, blurred_x, blurred_y;
int scale = 1;
int delta = 0;
int ddepth = CV_16S;
int kernSize = 3;
Sobel(gray, grad_x, ddepth, 1, 0, kernSize, scale, delta, BORDER_DEFAULT);
Sobel(gray, grad_y, ddepth, 0, 1, kernSize, scale, delta, BORDER_DEFAULT);
GaussianBlur(grad_x, blurred_x, Size(5, 5), 30);
GaussianBlur(grad_y, blurred_y, Size(5, 5), 30);
return { blurred_x, blurred_y };
}
array<Mat, 2> external(array<Mat, 2> &edgeMat, Mat &x, Mat &y) {
array<Mat, 2> ext;
ext[0] = { Size{ 1, POINTS_NUM }, CV_32FC1 };
ext[1] = { Size{ 1, POINTS_NUM }, CV_32FC1 };
for (size_t i = 0; i < POINTS_NUM; ++i) {
ext[0].at<float>(0, i) = - edgeMat[0].at<short>(y.at<float>(0, i), x.at<float>(0, i));
ext[1].at<float>(0, i) = - edgeMat[1].at<short>(y.at<float>(0, i), x.at<float>(0, i));
}
return ext;
}
As you can see, the contour points converge in a very strange way and not towards the edge of the lip (that was the result I would expect).
I am not able to understand if it is an error about implementation or about tuning the parameters or it is just is normal behaviour and I misunderstood something about the algorithm.
I have some doubts on the derivative matrices, I think that they should be regularized in some way, but I am not sure which is the right one. Can someone help me?
The only implementations I have found are of the greedy method.
I’m using a modified version of a gauss-newton method to refine a pose estimate using OpenCV. The unmodified code can be found here: http://people.rennes.inria.fr/Eric.Marchand/pose-estimation/tutorial-pose-gauss-newton-opencv.html
The details of this approach are outlined in the corresponding paper:
Marchand, Eric, Hideaki Uchiyama, and Fabien Spindler. "Pose
estimation for augmented reality: a hands-on survey." IEEE
transactions on visualization and computer graphics 22.12 (2016):
2633-2651.
A PDF can be found here: https://hal.inria.fr/hal-01246370/document
The part that is relevant (Pages 4 and 5) are screencapped below:
Here is what I have done. First, I’ve (hopefully) “corrected” some errors: (a) dt and dR can be passed by reference to exponential_map() (even though cv::Mat is essentially a pointer). (b) The last entry of each 2x6 Jacobian matrix, J.at<double>(i*2+1,5), was -x[i].y but should be -x[i].x. (c) I’ve also tried using a different formula for the projection. Specifically, one that includes the focal length and principal point:
xq.at<double>(i*2,0) = cx + fx * cX.at<double>(0,0) / cX.at<double>(2,0);
xq.at<double>(i*2+1,0) = cy + fy * cX.at<double>(1,0) / cX.at<double>(2,0);
Here is the relevant code I am using, in its entirety (control starts at optimizePose3()):
void exponential_map(const cv::Mat &v, cv::Mat &dt, cv::Mat &dR)
{
double vx = v.at<double>(0,0);
double vy = v.at<double>(1,0);
double vz = v.at<double>(2,0);
double vtux = v.at<double>(3,0);
double vtuy = v.at<double>(4,0);
double vtuz = v.at<double>(5,0);
cv::Mat tu = (cv::Mat_<double>(3,1) << vtux, vtuy, vtuz); // theta u
cv::Rodrigues(tu, dR);
double theta = sqrt(tu.dot(tu));
double sinc = (fabs(theta) < 1.0e-8) ? 1.0 : sin(theta) / theta;
double mcosc = (fabs(theta) < 2.5e-4) ? 0.5 : (1.-cos(theta)) / theta / theta;
double msinc = (fabs(theta) < 2.5e-4) ? (1./6.) : (1.-sin(theta)/theta) / theta / theta;
dt.at<double>(0,0) = vx*(sinc + vtux*vtux*msinc)
+ vy*(vtux*vtuy*msinc - vtuz*mcosc)
+ vz*(vtux*vtuz*msinc + vtuy*mcosc);
dt.at<double>(1,0) = vx*(vtux*vtuy*msinc + vtuz*mcosc)
+ vy*(sinc + vtuy*vtuy*msinc)
+ vz*(vtuy*vtuz*msinc - vtux*mcosc);
dt.at<double>(2,0) = vx*(vtux*vtuz*msinc - vtuy*mcosc)
+ vy*(vtuy*vtuz*msinc + vtux*mcosc)
+ vz*(sinc + vtuz*vtuz*msinc);
}
void optimizePose3(const PoseEstimation &pose,
std::vector<FeatureMatch> &feature_matches,
PoseEstimation &optimized_pose) {
//Set camera parameters
double fx = camera_matrix.at<double>(0, 0); //Focal length
double fy = camera_matrix.at<double>(1, 1);
double cx = camera_matrix.at<double>(0, 2); //Principal point
double cy = camera_matrix.at<double>(1, 2);
auto inlier_matches = getInliers(pose, feature_matches);
std::vector<cv::Point3d> wX;
std::vector<cv::Point2d> x;
const unsigned int npoints = inlier_matches.size();
cv::Mat J(2*npoints, 6, CV_64F);
double lambda = 0.25;
cv::Mat xq(npoints*2, 1, CV_64F);
cv::Mat xn(npoints*2, 1, CV_64F);
double residual=0, residual_prev;
cv::Mat Jp;
for(auto i = 0u; i < npoints; i++) {
//Model points
const cv::Point2d &M = inlier_matches[i].model_point();
wX.emplace_back(M.x, M.y, 0.0);
//Imaged points
const cv::Point2d &I = inlier_matches[i].image_point();
xn.at<double>(i*2,0) = I.x; // x
xn.at<double>(i*2+1,0) = I.y; // y
x.push_back(I);
}
//Initial estimation
cv::Mat cRw = pose.rotation_matrix;
cv::Mat ctw = pose.translation_vector;
int nIters = 0;
// Iterative Gauss-Newton minimization loop
do {
for (auto i = 0u; i < npoints; i++) {
cv::Mat cX = cRw * cv::Mat(wX[i]) + ctw; // Update cX, cY, cZ
// Update x(q)
//xq.at<double>(i*2,0) = cX.at<double>(0,0) / cX.at<double>(2,0); // x(q) = cX/cZ
//xq.at<double>(i*2+1,0) = cX.at<double>(1,0) / cX.at<double>(2,0); // y(q) = cY/cZ
xq.at<double>(i*2,0) = cx + fx * cX.at<double>(0,0) / cX.at<double>(2,0);
xq.at<double>(i*2+1,0) = cy + fy * cX.at<double>(1,0) / cX.at<double>(2,0);
// Update J using equation (11)
J.at<double>(i*2,0) = -1 / cX.at<double>(2,0); // -1/cZ
J.at<double>(i*2,1) = 0;
J.at<double>(i*2,2) = x[i].x / cX.at<double>(2,0); // x/cZ
J.at<double>(i*2,3) = x[i].x * x[i].y; // xy
J.at<double>(i*2,4) = -(1 + x[i].x * x[i].x); // -(1+x^2)
J.at<double>(i*2,5) = x[i].y; // y
J.at<double>(i*2+1,0) = 0;
J.at<double>(i*2+1,1) = -1 / cX.at<double>(2,0); // -1/cZ
J.at<double>(i*2+1,2) = x[i].y / cX.at<double>(2,0); // y/cZ
J.at<double>(i*2+1,3) = 1 + x[i].y * x[i].y; // 1+y^2
J.at<double>(i*2+1,4) = -x[i].x * x[i].y; // -xy
J.at<double>(i*2+1,5) = -x[i].x; // -x
}
cv::Mat e_q = xq - xn; // Equation (7)
cv::Mat Jp = J.inv(cv::DECOMP_SVD); // Compute pseudo inverse of the Jacobian
cv::Mat dq = -lambda * Jp * e_q; // Equation (10)
cv::Mat dctw(3, 1, CV_64F), dcRw(3, 3, CV_64F);
exponential_map(dq, dctw, dcRw);
cRw = dcRw.t() * cRw; // Update the pose
ctw = dcRw.t() * (ctw - dctw);
residual_prev = residual; // Memorize previous residual
residual = e_q.dot(e_q); // Compute the actual residual
std::cout << "residual_prev: " << residual_prev << std::endl;
std::cout << "residual: " << residual << std::endl << std::endl;
nIters++;
} while (fabs(residual - residual_prev) > 0);
//} while (nIters < 30);
optimized_pose.rotation_matrix = cRw;
optimized_pose.translation_vector = ctw;
cv::Rodrigues(optimized_pose.rotation_matrix, optimized_pose.rotation_vector);
}
Even when I use the functions as given, it does not produce the correct results. My initial pose estimate is very close to optimal, but when I try run the program, the method takes a very long time to converge - and when it does, the results are very wrong. I’m not sure what could be wrong and I’m out of ideas. I’m confident my inliers are actually inliers (they were chosen using an M-estimator). I’ve compared the results from exponential map with those from other implementations, and they seem to agree.
So, where is the error in this gauss-newton implementation for pose optimization? I’ve tried to make things as easy as possible for anyone willing to lend a hand. Let me know if there is anymore information I can provide. Any help would be greatly appreciated. Thanks.
Edit: 2019/05/13
There is now solvePnPRefineVVS function in OpenCV.
Also, you should use x and y calculated from the current estimated pose instead.
In the cited paper, they expressed the measurements x in the normalized camera frame (at z=1).
When working with real data, you have:
(u,v): 2D image coordinates (e.g. keypoints, corner locations, etc.)
K: the intrinsic parameters (obtained after calibrating the camera)
D: the distortion coefficients (obtained after calibrating the camera)
To compute the 2D image coordinates in the normalized camera frame, you can use in OpenCV the function cv::undistortPoints() (link to my answer about cv::projectPoints() and cv::undistortPoints()).
When there is no distortion, the computation (also called "reverse perspective transformation") is:
x = (u - cx) / fx
y = (v - cy) / fy
I'm trying perspective transformation of an image using a homography matrix.
Given translation and rotation, I made a homography matrix and applied it to the perspective transformation as:
Mat srcImg = imread("tests/image3.jp2", IMREAD_COLOR);
Mat dstImg, H;
Get_Homography(H, srcImg.size());
warpPerspective(srcImg, dstImg, H, dstImg.size());
imshow("output", dstImg);
and
#define FL_x 1000.0
#define FL_y 1000.0
void Get_Homography(Mat &H_out, cv::Size size)
{
static float H_uc[9], C[9], C_inv[9], H[9], C_inv_H_uc[9];
static float H33[3][3];
static float R[9];
static float T[3];
static float n[3];
static float d;
static float nTd[9];
static float phi, the, psi;
n[0] = n[1] = 0.0;
n[2] = -1.0;
T[0] = -500;
T[1] = -500;
T[2] = 0.0;
d = 100.0;
phi = 0.0*D2R;
the = 0.0*D2R;
psi = 0.0*D2R;
matMult(T, n, 3, 1, 3, nTd);
matMult(nTd, &d, 9, 1, 1, nTd);
getDCM_I2B(phi, the, psi, R);
matAdd(R, nTd, 3, 3, H_uc);
C[0] = FL_x; C[1] = 0.0; C[2] = size.width / 2.0;
C[3] = 0.0; C[4] = FL_y; C[5] = size.height / 2.0;
C[6] = 0.0; C[7] = 0.0; C[8] = 1.0;
matInv33(C, C_inv);
matMult(C_inv, H_uc, 3, 3, 3, C_inv_H_uc);
matMult(C_inv_H_uc, C, 3, 3, 3, H);
H33[0][0] = H[0];
H33[0][1] = H[1];
H33[0][2] = H[2];
H33[1][0] = H[3];
H33[1][1] = H[4];
H33[1][2] = H[5];
H33[2][0] = H[6];
H33[2][1] = H[7];
H33[2][2] = H[8];
H_out = Mat(3, 3, CV_32F, H33);
return;
}
The rotations by z-axis (any value at "psi") work alright.
But when I put any value (even 1.0 deg) to "the" or "phi", the resultant image is awkward. I cannot recognize what it is.
And when I put [-500, -500, 0] to T (translation), it produces a shifted image as if it is taken in a different position (shifted in right direction), but I think -500 -500 are too big. For d = 1.0, the resultant image just shows a few pixel shift (not recognizably).
I thought my implementation of constructing a homography matrix is right, but the results are awkward.
Could you give me some advice on this?
Thank you.
I'd like to rotate an image, but I can't obtain the rotated image without cropping
My original image:
Now I use this code:
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
// Compile with g++ code.cpp -lopencv_core -lopencv_highgui -lopencv_imgproc
int main()
{
cv::Mat src = cv::imread("im.png", CV_LOAD_IMAGE_UNCHANGED);
cv::Mat dst;
cv::Point2f pc(src.cols/2., src.rows/2.);
cv::Mat r = cv::getRotationMatrix2D(pc, -45, 1.0);
cv::warpAffine(src, dst, r, src.size()); // what size I should use?
cv::imwrite("rotated_im.png", dst);
return 0;
}
And obtain the following image:
But I'd like to obtain this:
My answer is inspired by the following posts / blog entries:
Rotate cv::Mat using cv::warpAffine offsets destination image
http://john.freml.in/opencv-rotation
Main ideas:
Adjusting the rotation matrix by adding a translation to the new image center
Using cv::RotatedRect to rely on existing opencv functionality as much as possible
Code tested with opencv 3.4.1:
#include "opencv2/opencv.hpp"
int main()
{
cv::Mat src = cv::imread("im.png", CV_LOAD_IMAGE_UNCHANGED);
double angle = -45;
// get rotation matrix for rotating the image around its center in pixel coordinates
cv::Point2f center((src.cols-1)/2.0, (src.rows-1)/2.0);
cv::Mat rot = cv::getRotationMatrix2D(center, angle, 1.0);
// determine bounding rectangle, center not relevant
cv::Rect2f bbox = cv::RotatedRect(cv::Point2f(), src.size(), angle).boundingRect2f();
// adjust transformation matrix
rot.at<double>(0,2) += bbox.width/2.0 - src.cols/2.0;
rot.at<double>(1,2) += bbox.height/2.0 - src.rows/2.0;
cv::Mat dst;
cv::warpAffine(src, dst, rot, bbox.size());
cv::imwrite("rotated_im.png", dst);
return 0;
}
Just try the code below, the idea is simple:
You need to create a blank image with the maximum size you're expecting while rotating at any angle. Here you should use Pythagoras as mentioned in the above comments.
Now copy the source image to the newly created image and pass it to warpAffine. Here you should use the centre of newly created image for rotation.
After warpAffine if you need to crop exact image for this translate four corners of source image in enlarged image using rotation matrix as described here
Find minimum x and minimum y for top corner, and maximum x and maximum y for bottom corner from the above result to crop image.
This is the code:
int theta = 0;
Mat src,frame, frameRotated;
src = imread("rotate.png",1);
cout<<endl<<endl<<"Press '+' to rotate anti-clockwise and '-' for clockwise 's' to save" <<endl<<endl;
int diagonal = (int)sqrt(src.cols*src.cols+src.rows*src.rows);
int newWidth = diagonal;
int newHeight =diagonal;
int offsetX = (newWidth - src.cols) / 2;
int offsetY = (newHeight - src.rows) / 2;
Mat targetMat(newWidth, newHeight, src.type());
Point2f src_center(targetMat.cols/2.0F, targetMat.rows/2.0F);
while(1){
src.copyTo(frame);
double radians = theta * M_PI / 180.0;
double sin = abs(std::sin(radians));
double cos = abs(std::cos(radians));
frame.copyTo(targetMat.rowRange(offsetY, offsetY + frame.rows).colRange(offsetX, offsetX + frame.cols));
Mat rot_mat = getRotationMatrix2D(src_center, theta, 1.0);
warpAffine(targetMat, frameRotated, rot_mat, targetMat.size());
//Calculate bounding rect and for exact image
//Reference:- https://stackoverflow.com/questions/19830477/find-the-bounding-rectangle-of-rotated-rectangle/19830964?noredirect=1#19830964
Rect bound_Rect(frame.cols,frame.rows,0,0);
int x1 = offsetX;
int x2 = offsetX+frame.cols;
int x3 = offsetX;
int x4 = offsetX+frame.cols;
int y1 = offsetY;
int y2 = offsetY;
int y3 = offsetY+frame.rows;
int y4 = offsetY+frame.rows;
Mat co_Ordinate = (Mat_<double>(3,4) << x1, x2, x3, x4,
y1, y2, y3, y4,
1, 1, 1, 1 );
Mat RotCo_Ordinate = rot_mat * co_Ordinate;
for(int i=0;i<4;i++){
if(RotCo_Ordinate.at<double>(0,i)<bound_Rect.x)
bound_Rect.x=(int)RotCo_Ordinate.at<double>(0,i); //access smallest
if(RotCo_Ordinate.at<double>(1,i)<bound_Rect.y)
bound_Rect.y=RotCo_Ordinate.at<double>(1,i); //access smallest y
}
for(int i=0;i<4;i++){
if(RotCo_Ordinate.at<double>(0,i)>bound_Rect.width)
bound_Rect.width=(int)RotCo_Ordinate.at<double>(0,i); //access largest x
if(RotCo_Ordinate.at<double>(1,i)>bound_Rect.height)
bound_Rect.height=RotCo_Ordinate.at<double>(1,i); //access largest y
}
bound_Rect.width=bound_Rect.width-bound_Rect.x;
bound_Rect.height=bound_Rect.height-bound_Rect.y;
Mat cropedResult;
Mat ROI = frameRotated(bound_Rect);
ROI.copyTo(cropedResult);
imshow("Result", cropedResult);
imshow("frame", frame);
imshow("rotated frame", frameRotated);
char k=waitKey();
if(k=='+') theta+=10;
if(k=='-') theta-=10;
if(k=='s') imwrite("rotated.jpg",cropedResult);
if(k==27) break;
}
Cropped Image
Thanks Robula!
Actually, you do not need to compute sine and cosine twice.
import cv2
def rotate_image(mat, angle):
# angle in degrees
height, width = mat.shape[:2]
image_center = (width/2, height/2)
rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.)
abs_cos = abs(rotation_mat[0,0])
abs_sin = abs(rotation_mat[0,1])
bound_w = int(height * abs_sin + width * abs_cos)
bound_h = int(height * abs_cos + width * abs_sin)
rotation_mat[0, 2] += bound_w/2 - image_center[0]
rotation_mat[1, 2] += bound_h/2 - image_center[1]
rotated_mat = cv2.warpAffine(mat, rotation_mat, (bound_w, bound_h))
return rotated_mat
Thanks #Haris! Here's the Python version:
def rotate_image(image, angle):
'''Rotate image "angle" degrees.
How it works:
- Creates a blank image that fits any rotation of the image. To achieve
this, set the height and width to be the image's diagonal.
- Copy the original image to the center of this blank image
- Rotate using warpAffine, using the newly created image's center
(the enlarged blank image center)
- Translate the four corners of the source image in the enlarged image
using homogenous multiplication of the rotation matrix.
- Crop the image according to these transformed corners
'''
diagonal = int(math.sqrt(pow(image.shape[0], 2) + pow(image.shape[1], 2)))
offset_x = (diagonal - image.shape[0])/2
offset_y = (diagonal - image.shape[1])/2
dst_image = np.zeros((diagonal, diagonal, 3), dtype='uint8')
image_center = (diagonal/2, diagonal/2)
R = cv2.getRotationMatrix2D(image_center, angle, 1.0)
dst_image[offset_x:(offset_x + image.shape[0]), \
offset_y:(offset_y + image.shape[1]), \
:] = image
dst_image = cv2.warpAffine(dst_image, R, (diagonal, diagonal), flags=cv2.INTER_LINEAR)
# Calculate the rotated bounding rect
x0 = offset_x
x1 = offset_x + image.shape[0]
x2 = offset_x
x3 = offset_x + image.shape[0]
y0 = offset_y
y1 = offset_y
y2 = offset_y + image.shape[1]
y3 = offset_y + image.shape[1]
corners = np.zeros((3,4))
corners[0,0] = x0
corners[0,1] = x1
corners[0,2] = x2
corners[0,3] = x3
corners[1,0] = y0
corners[1,1] = y1
corners[1,2] = y2
corners[1,3] = y3
corners[2:] = 1
c = np.dot(R, corners)
x = int(c[0,0])
y = int(c[1,0])
left = x
right = x
up = y
down = y
for i in range(4):
x = int(c[0,i])
y = int(c[1,i])
if (x < left): left = x
if (x > right): right = x
if (y < up): up = y
if (y > down): down = y
h = down - up
w = right - left
cropped = np.zeros((w, h, 3), dtype='uint8')
cropped[:, :, :] = dst_image[left:(left+w), up:(up+h), :]
return cropped
Increase the image canvas (equally from the center without changing the image size) so that it can fit the image after rotation, then apply warpAffine:
Mat img = imread ("/path/to/image", 1);
double offsetX, offsetY;
double angle = -45;
double width = img.size().width;
double height = img.size().height;
Point2d center = Point2d (width / 2, height / 2);
Rect bounds = RotatedRect (center, img.size(), angle).boundingRect();
Mat resized = Mat::zeros (bounds.size(), img.type());
offsetX = (bounds.width - width) / 2;
offsetY = (bounds.height - height) / 2;
Rect roi = Rect (offsetX, offsetY, width, height);
img.copyTo (resized (roi));
center += Point2d (offsetX, offsetY);
Mat M = getRotationMatrix2D (center, angle, 1.0);
warpAffine (resized, resized, M, resized.size());
After searching around for a clean and easy to understand solution and reading through the answers above trying to understand them, I eventually came up with a solution using trigonometry.
I hope this helps somebody :)
import cv2
import math
def rotate_image(mat, angle):
height, width = mat.shape[:2]
image_center = (width / 2, height / 2)
rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1)
radians = math.radians(angle)
sin = math.sin(radians)
cos = math.cos(radians)
bound_w = int((height * abs(sin)) + (width * abs(cos)))
bound_h = int((height * abs(cos)) + (width * abs(sin)))
rotation_mat[0, 2] += ((bound_w / 2) - image_center[0])
rotation_mat[1, 2] += ((bound_h / 2) - image_center[1])
rotated_mat = cv2.warpAffine(mat, rotation_mat, (bound_w, bound_h))
return rotated_mat
EDIT: Please refer to #Remi Cuingnet's answer below.
A python version of rotating an image and take control of the padded black coloured region you can use the scipy.ndimage.rotate. Here is an example:
from skimage import io
from scipy import ndimage
image = io.imread('https://www.pyimagesearch.com/wp-
content/uploads/2019/12/tensorflow2_install_ubuntu_header.jpg')
io.imshow(image)
plt.show()
rotated = ndimage.rotate(image, angle=234, mode='nearest')
rotated = cv2.resize(rotated, (image.shape[:2]))
# rotated = cv2.cvtColor(rotated, cv2.COLOR_BGR2RGB)
# cv2.imwrite('rotated.jpg', rotated)
io.imshow(rotated)
plt.show()
If you have a rotation and a scaling of the image:
#include "opencv2/opencv.hpp"
#include <functional>
#include <vector>
bool compareCoords(cv::Point2f p1, cv::Point2f p2, char coord)
{
assert(coord == 'x' || coord == 'y');
if (coord == 'x')
return p1.x < p2.x;
return p1.y < p2.y;
}
int main(int argc, char** argv)
{
cv::Mat image = cv::imread("lenna.png");
float angle = 45.0; // degrees
float scale = 0.5;
cv::Mat_<float> rot_mat = cv::getRotationMatrix2D( cv::Point2f( 0.0f, 0.0f ), angle, scale );
// Image corners
cv::Point2f pA = cv::Point2f(0.0f, 0.0f);
cv::Point2f pB = cv::Point2f(image.cols, 0.0f);
cv::Point2f pC = cv::Point2f(image.cols, image.rows);
cv::Point2f pD = cv::Point2f(0.0f, image.rows);
std::vector<cv::Point2f> pts = { pA, pB, pC, pD };
std::vector<cv::Point2f> ptsTransf;
cv::transform(pts, ptsTransf, rot_mat );
using namespace std::placeholders;
float minX = std::min_element(ptsTransf.begin(), ptsTransf.end(), std::bind(compareCoords, _1, _2, 'x'))->x;
float maxX = std::max_element(ptsTransf.begin(), ptsTransf.end(), std::bind(compareCoords, _1, _2, 'x'))->x;
float minY = std::min_element(ptsTransf.begin(), ptsTransf.end(), std::bind(compareCoords, _1, _2, 'y'))->y;
float maxY = std::max_element(ptsTransf.begin(), ptsTransf.end(), std::bind(compareCoords, _1, _2, 'y'))->y;
float newW = maxX - minX;
float newH = maxY - minY;
cv::Mat_<float> trans_mat = (cv::Mat_<float>(2,3) << 0, 0, -minX, 0, 0, -minY);
cv::Mat_<float> M = rot_mat + trans_mat;
cv::Mat warpedImage;
cv::warpAffine( image, warpedImage, M, cv::Size(newW, newH) );
cv::imshow("lenna", image);
cv::imshow("Warped lenna", warpedImage);
cv::waitKey();
cv::destroyAllWindows();
return 0;
}
Thanks to everyone for this post, it has been super useful. However, I have found some black lines left and up (using Rose's python version) when rotating 90º. The problem seemed to be some int() roundings. In addition to that, I have changed the sign of the angle to make it grow clockwise.
def rotate_image(image, angle):
'''Rotate image "angle" degrees.
How it works:
- Creates a blank image that fits any rotation of the image. To achieve
this, set the height and width to be the image's diagonal.
- Copy the original image to the center of this blank image
- Rotate using warpAffine, using the newly created image's center
(the enlarged blank image center)
- Translate the four corners of the source image in the enlarged image
using homogenous multiplication of the rotation matrix.
- Crop the image according to these transformed corners
'''
diagonal = int(math.ceil(math.sqrt(pow(image.shape[0], 2) + pow(image.shape[1], 2))))
offset_x = (diagonal - image.shape[0])/2
offset_y = (diagonal - image.shape[1])/2
dst_image = np.zeros((diagonal, diagonal, 3), dtype='uint8')
image_center = (float(diagonal-1)/2, float(diagonal-1)/2)
R = cv2.getRotationMatrix2D(image_center, -angle, 1.0)
dst_image[offset_x:(offset_x + image.shape[0]), offset_y:(offset_y + image.shape[1]), :] = image
dst_image = cv2.warpAffine(dst_image, R, (diagonal, diagonal), flags=cv2.INTER_LINEAR)
# Calculate the rotated bounding rect
x0 = offset_x
x1 = offset_x + image.shape[0]
x2 = offset_x + image.shape[0]
x3 = offset_x
y0 = offset_y
y1 = offset_y
y2 = offset_y + image.shape[1]
y3 = offset_y + image.shape[1]
corners = np.zeros((3,4))
corners[0,0] = x0
corners[0,1] = x1
corners[0,2] = x2
corners[0,3] = x3
corners[1,0] = y0
corners[1,1] = y1
corners[1,2] = y2
corners[1,3] = y3
corners[2:] = 1
c = np.dot(R, corners)
x = int(round(c[0,0]))
y = int(round(c[1,0]))
left = x
right = x
up = y
down = y
for i in range(4):
x = c[0,i]
y = c[1,i]
if (x < left): left = x
if (x > right): right = x
if (y < up): up = y
if (y > down): down = y
h = int(round(down - up))
w = int(round(right - left))
left = int(round(left))
up = int(round(up))
cropped = np.zeros((w, h, 3), dtype='uint8')
cropped[:, :, :] = dst_image[left:(left+w), up:(up+h), :]
return cropped
Go version (using gocv) of #robula and #remi-cuingnet
func rotateImage(mat *gocv.Mat, angle float64) *gocv.Mat {
height := mat.Rows()
width := mat.Cols()
imgCenter := image.Point{X: width/2, Y: height/2}
rotationMat := gocv.GetRotationMatrix2D(imgCenter, -angle, 1.0)
absCos := math.Abs(rotationMat.GetDoubleAt(0, 0))
absSin := math.Abs(rotationMat.GetDoubleAt(0, 1))
boundW := float64(height) * absSin + float64(width) * absCos
boundH := float64(height) * absCos + float64(width) * absSin
rotationMat.SetDoubleAt(0, 2, rotationMat.GetDoubleAt(0, 2) + (boundW / 2) - float64(imgCenter.X))
rotationMat.SetDoubleAt(1, 2, rotationMat.GetDoubleAt(1, 2) + (boundH / 2) - float64(imgCenter.Y))
gocv.WarpAffine(*mat, mat, rotationMat, image.Point{ X: int(boundW), Y: int(boundH) })
return mat
}
I rotate in the same matrice in-memory, make a new matrice if you don't want to alter it
For anyone using Emgu.CV or OpenCvSharp wrapper in .NET, there's a C# implement of Lars Schillingmann's answer:
Emgu.CV:
using Emgu.CV;
using Emgu.CV.CvEnum;
using Emgu.CV.Structure;
public static class MatExtension
{
/// <summary>
/// <see>https://stackoverflow.com/questions/22041699/rotate-an-image-without-cropping-in-opencv-in-c/75451191#75451191</see>
/// </summary>
public static Mat Rotate(this Mat src, float degrees)
{
degrees = -degrees; // counter-clockwise to clockwise
var center = new PointF((src.Width - 1) / 2f, (src.Height - 1) / 2f);
var rotationMat = new Mat();
CvInvoke.GetRotationMatrix2D(center, degrees, 1, rotationMat);
var boundingRect = new RotatedRect(new(), src.Size, degrees).MinAreaRect();
rotationMat.Set(0, 2, rotationMat.Get<double>(0, 2) + (boundingRect.Width / 2f) - (src.Width / 2f));
rotationMat.Set(1, 2, rotationMat.Get<double>(1, 2) + (boundingRect.Height / 2f) - (src.Height / 2f));
var rotatedSrc = new Mat();
CvInvoke.WarpAffine(src, rotatedSrc, rotationMat, boundingRect.Size);
return rotatedSrc;
}
/// <summary>
/// <see>https://stackoverflow.com/questions/32255440/how-can-i-get-and-set-pixel-values-of-an-emgucv-mat-image/69537504#69537504</see>
/// </summary>
public static unsafe void Set<T>(this Mat mat, int row, int col, T value) where T : struct =>
_ = new Span<T>(mat.DataPointer.ToPointer(), mat.Rows * mat.Cols * mat.ElementSize)
{
[(row * mat.Cols) + col] = value
};
public static unsafe T Get<T>(this Mat mat, int row, int col) where T : struct =>
new ReadOnlySpan<T>(mat.DataPointer.ToPointer(), mat.Rows * mat.Cols * mat.ElementSize)
[(row * mat.Cols) + col];
}
OpenCvSharp:
OpenCvSharp already has Mat.Set<> method that functions same as mat.at<> in the original OpenCV, so we don't have to copy these methods from How can I get and set pixel values of an EmguCV Mat image?
using OpenCvSharp;
public static class MatExtension
{
/// <summary>
/// <see>https://stackoverflow.com/questions/22041699/rotate-an-image-without-cropping-in-opencv-in-c/75451191#75451191</see>
/// </summary>
public static Mat Rotate(this Mat src, float degrees)
{
degrees = -degrees; // counter-clockwise to clockwise
var center = new Point2f((src.Width - 1) / 2f, (src.Height - 1) / 2f);
var rotationMat = Cv2.GetRotationMatrix2D(center, degrees, 1);
var boundingRect = new RotatedRect(new(), new Size2f(src.Width, src.Height), degrees).BoundingRect();
rotationMat.Set(0, 2, rotationMat.Get<double>(0, 2) + (boundingRect.Width / 2f) - (src.Width / 2f));
rotationMat.Set(1, 2, rotationMat.Get<double>(1, 2) + (boundingRect.Height / 2f) - (src.Height / 2f));
var rotatedSrc = new Mat();
Cv2.WarpAffine(src, rotatedSrc, rotationMat, boundingRect.Size);
return rotatedSrc;
}
}
Also, you may want to mutate the src param instead of returning a new clone of it during rotation, for that you can just set the det param of WrapAffine() as the same with src: c++, opencv: Is it safe to use the same Mat for both source and destination images in filtering operation?
CvInvoke.WarpAffine(src, src, rotationMat, boundingRect.Size);
This is being called as in-place mode: https://answers.opencv.org/question/24/do-all-opencv-functions-support-in-place-mode-for-their-arguments/
Can the OpenCV function cvtColor be used to convert a matrix in place?
If it is just to rotate 90 degrees, maybe this code could be useful.
Mat img = imread("images.jpg");
Mat rt(img.rows, img.rows, CV_8U);
Point2f pc(img.cols / 2.0, img.rows / 2.0);
Mat r = getRotationMatrix2D(pc, 90, 1);
warpAffine(img, rt, r, rt.size());
imshow("rotated", rt);
Hope it's useful.
By the way, for 90º rotations only, here is a more efficient + accurate function:
def rotate_image_90(image, angle):
angle = -angle
rotated_image = image
if angle == 0:
pass
elif angle == 90:
rotated_image = np.rot90(rotated_image)
elif angle == 180 or angle == -180:
rotated_image = np.rot90(rotated_image)
rotated_image = np.rot90(rotated_image)
elif angle == -90:
rotated_image = np.rot90(rotated_image)
rotated_image = np.rot90(rotated_image)
rotated_image = np.rot90(rotated_image)
return rotated_image
If I have a texture, is it then possible to generate a normal-map for this texture, so it can be used for bump-mapping?
Or how are normal maps usually made?
Yes. Well, sort of. Normal maps can be accurately made from height-maps. Generally, you can also put a regular texture through and get decent results as well. Keep in mind there are other methods of making a normal map, such as taking a high-resolution model, making it low resolution, then doing ray casting to see what the normal should be for the low-resolution model to simulate the higher one.
For height-map to normal-map, you can use the Sobel Operator. This operator can be run in the x-direction, telling you the x-component of the normal, and then the y-direction, telling you the y-component. You can calculate z with 1.0 / strength where strength is the emphasis or "deepness" of the normal map. Then, take that x, y, and z, throw them into a vector, normalize it, and you have your normal at that point. Encode it into the pixel and you're done.
Here's some older incomplete-code that demonstrates this:
// pretend types, something like this
struct pixel
{
uint8_t red;
uint8_t green;
uint8_t blue;
};
struct vector3d; // a 3-vector with doubles
struct texture; // a 2d array of pixels
// determine intensity of pixel, from 0 - 1
const double intensity(const pixel& pPixel)
{
const double r = static_cast<double>(pPixel.red);
const double g = static_cast<double>(pPixel.green);
const double b = static_cast<double>(pPixel.blue);
const double average = (r + g + b) / 3.0;
return average / 255.0;
}
const int clamp(int pX, int pMax)
{
if (pX > pMax)
{
return pMax;
}
else if (pX < 0)
{
return 0;
}
else
{
return pX;
}
}
// transform -1 - 1 to 0 - 255
const uint8_t map_component(double pX)
{
return (pX + 1.0) * (255.0 / 2.0);
}
texture normal_from_height(const texture& pTexture, double pStrength = 2.0)
{
// assume square texture, not necessarily true in real code
texture result(pTexture.size(), pTexture.size());
const int textureSize = static_cast<int>(pTexture.size());
for (size_t row = 0; row < textureSize; ++row)
{
for (size_t column = 0; column < textureSize; ++column)
{
// surrounding pixels
const pixel topLeft = pTexture(clamp(row - 1, textureSize), clamp(column - 1, textureSize));
const pixel top = pTexture(clamp(row - 1, textureSize), clamp(column, textureSize));
const pixel topRight = pTexture(clamp(row - 1, textureSize), clamp(column + 1, textureSize));
const pixel right = pTexture(clamp(row, textureSize), clamp(column + 1, textureSize));
const pixel bottomRight = pTexture(clamp(row + 1, textureSize), clamp(column + 1, textureSize));
const pixel bottom = pTexture(clamp(row + 1, textureSize), clamp(column, textureSize));
const pixel bottomLeft = pTexture(clamp(row + 1, textureSize), clamp(column - 1, textureSize));
const pixel left = pTexture(clamp(row, textureSize), clamp(column - 1, textureSize));
// their intensities
const double tl = intensity(topLeft);
const double t = intensity(top);
const double tr = intensity(topRight);
const double r = intensity(right);
const double br = intensity(bottomRight);
const double b = intensity(bottom);
const double bl = intensity(bottomLeft);
const double l = intensity(left);
// sobel filter
const double dX = (tr + 2.0 * r + br) - (tl + 2.0 * l + bl);
const double dY = (bl + 2.0 * b + br) - (tl + 2.0 * t + tr);
const double dZ = 1.0 / pStrength;
math::vector3d v(dX, dY, dZ);
v.normalize();
// convert to rgb
result(row, column) = pixel(map_component(v.x), map_component(v.y), map_component(v.z));
}
}
return result;
}
There's probably many ways to generate a Normal map, but like others said, you can do it from a Height Map, and 3d packages like XSI/3dsmax/Blender/any of them can output one for you as an image.
You can then output and RGB image with the Nvidia plugin for photoshop, an algorithm to convert it or you might be able to output it directly from those 3d packages with 3rd party plugins.
Be aware that in some case, you might need to invert channels (R, G or B) from the generated normal map.
Here's some resources link with examples and more complete explanation:
http://developer.nvidia.com/object/photoshop_dds_plugins.html
http://en.wikipedia.org/wiki/Normal_mapping
http://www.vrgeo.org/fileadmin/VRGeo/Bilder/VRGeo_Papers/jgt2002normalmaps.pdf
I don't think normal maps are generated from a texture. they are generated from a model.
just as texturing allows you to define complex colour detail with minimal polys (as opposed to just using millions of ploys and just vertex colours to define the colour on your mesh)
A normal map allows you to define complex normal detail with minimal polys.
I believe normal maps are usually generated from a higher res mesh, and then is used with a low res mesh.
I'm sure 3D tools, such as 3ds max or maya, as well as more specific tools will do this for you. unlike textures, I don't think they are usually done by hand.
but they are generated from the mesh, not the texture.
I suggest starting with OpenCV, due to its richness in algorithms. Here's one I wrote that iteratively blurs the normal map and weights those to the overall value, essentially creating more of a topological map.
#define ROW_PTR(img, y) ((uchar*)((img).data + (img).step * y))
cv::Mat normalMap(const cv::Mat& bwTexture, double pStrength)
{
// assume square texture, not necessarily true in real code
int scale = 1.0;
int delta = 127;
cv::Mat sobelZ, sobelX, sobelY;
cv::Sobel(bwTexture, sobelX, CV_8U, 1, 0, 13, scale, delta, cv::BORDER_DEFAULT);
cv::Sobel(bwTexture, sobelY, CV_8U, 0, 1, 13, scale, delta, cv::BORDER_DEFAULT);
sobelZ = cv::Mat(bwTexture.rows, bwTexture.cols, CV_8UC1);
for(int y=0; y<bwTexture.rows; y++) {
const uchar *sobelXPtr = ROW_PTR(sobelX, y);
const uchar *sobelYPtr = ROW_PTR(sobelY, y);
uchar *sobelZPtr = ROW_PTR(sobelZ, y);
for(int x=0; x<bwTexture.cols; x++) {
double Gx = double(sobelXPtr[x]) / 255.0;
double Gy = double(sobelYPtr[x]) / 255.0;
double Gz = pStrength * sqrt(Gx * Gx + Gy * Gy);
uchar value = uchar(Gz * 255.0);
sobelZPtr[x] = value;
}
}
std::vector<cv::Mat>planes;
planes.push_back(sobelX);
planes.push_back(sobelY);
planes.push_back(sobelZ);
cv::Mat normalMap;
cv::merge(planes, normalMap);
cv::Mat originalNormalMap = normalMap.clone();
cv::Mat normalMapBlurred;
for (int i=0; i<3; i++) {
cv::GaussianBlur(normalMap, normalMapBlurred, cv::Size(13, 13), 5, 5);
addWeighted(normalMap, 0.4, normalMapBlurred, 0.6, 0, normalMap);
}
addWeighted(originalNormalMap, 0.3, normalMapBlurred, 0.7, 0, normalMap);
return normalMap;
}