IMAGE
I am having source image on the left. I distorted this image manually using grids from paint and result is on the right side. I was wondering if opencv offers any function which compares both images and gives some transformation matrix which I can use for further use. e.g. Passing this matrix to the video and I have distorted video like the image on right side.
I appreciate the help of this community :)
I was wondering if opencv offers any function which compares both images and gives some transformation matrix which I can use for further use.
Neither OpenCV nor any other library can calculate transformation matrix for non-linear transformation. It is impossible to represent non-linear transformation as a transformation matrix. However, one may find a transformation matrix that will be linear approximation of non-linear transformation. Such approximations should not necessarily yield satisfactory results.
As the image that you distort has a check pattern, you may describe the distortion by the movement of the corners of squares. I suggest that you try the following two-step approach:
Step1: Find corners both in original and distorted images.
OpenCV has a bunch of functions that are used in camera calibration and are designed for check pattern images. The following function uses cv2.findChessboardCorners function from OpenCV to find the corner points. Note: get_chessboard_corner_points returns relative coordinates, not the exact ones.
def get_chessboard_corner_points(img, corner_shape=(10,15)):
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 30, 0.01)
success, corners = cv2.findChessboardCorners(img, corner_shape)
if success:
corners = cv2.cornerSubPix(img, corners, (11,11), (-1,-1), criteria)
corners = np.squeeze(corners, axis=1)
for i in range(len(corners)):
corners[i] = (corners[i][0]/img.shape[1], corners[i][1]/img.shape[0])
#You may also try adding boundary points of the image by uncommenting the next line.
#corners = np.append(corners, np.array([[0,0],[0,1],[1,0],[1,1]]), 0)
#In addition you may also add other boundary points acording to your needs
return corners
return []
Step2: Distort each frame of video using interpolation and points found in the previous step.
Now when you have the points you need to write the distortion part.The distort_image function will do the job, though you may change it according to your needs. Note: src and dest contain relative coordinates. For target image we calculate and keep the exact coordinates in src_exact and dest_exact respectively. Then, griddata from SciPy is being used for interpolation.
def distort_image(img, src, dest):
height, width = img.shape[:2]
src_exact = []
dest_exact = []
for i in range(len(src)):
src_exact.append((int(src[i][1]*height), int(src[i][0]*width)))
dest_exact.append((int(dest[i][1]*height), int(dest[i][0]*width)))
grid_x, grid_y = np.mgrid[0:height - 1:complex(0, height), 0:width - 1:complex(0, width)]
grid_z = griddata(dest_exact, src_exact, (grid_x, grid_y), method='cubic')
map_x = np.append([], [ar[:, 1] for ar in grid_z]).reshape(height, width)
map_y = np.append([], [ar[:, 0] for ar in grid_z]).reshape(height, width)
map_x_32 = map_x.astype('float32')
map_y_32 = map_y.astype('float32')
distorted = cv2.remap(img, map_x_32, map_y_32, cv2.INTER_CUBIC)
return distorted
And here are a few lines that allow you to test the above presented functions.
import cv2
import numpy as np
from scipy.interpolate import griddata
TEST_IMG_PATH = "" #path to original grid
SOURCE_IMG_PATH = "" #path to distorted grid
DEST_IMG_PATH = "" #path to image to be distorted(this can be replaced by a video frame)
CORNER_SHAPE = (10, 15)
src_img = cv2.cvtColor(cv2.imread(SOURCE_IMG_PATH), cv2.COLOR_BGR2GRAY)
dest_img = cv2.cvtColor(cv2.imread(DEST_IMG_PATH), cv2.COLOR_BGR2GRAY)
test = cv2.cvtColor(cv2.imread(TEST_IMG_PATH), cv2.COLOR_BGR2RGB)
src_points = get_chessboard_corner_points(src_img, CORNER_SHAPE)
dest_points = get_chessboard_corner_points(dest_img, CORNER_SHAPE)
if len(src_points) != 0 and len(dest_points) != 0:
result = distort_image(test, src_points, dest_points)
Related
I'm attempting to be able to triangulate a 3D position using the OpenCV python library by calibrating two cameras and using the stereoCalibrate function. I want the cameras to be able to be attached to the ends of a bar and measure positions around 5m away from the cameras. The majority of code is similar to the code used by Temuge Batpurev (Source: https://temugeb.github.io/opencv/python/2021/02/02/stereo-camera-calibration-and-triangulation.html) and when I run his calibration images through my code, the RSME value ("ret" in his code, retStereo in my code) returned is the 2.4 expected as Temuge outlined.
When I use my own images, which are time-synched through GPS on GoPro 10s, I average around 50. Originally, I thought it was because I was calibrating for too large of an area but close up I still get around 50
Current Calibration Method:
Starting close to the cameras, without the chessboard going out of frame for either camera, slowly waving the board around. Ensuring it runs along the edges of the frames for both cameras as well as the center before moving backwards and repeating this at various depths.
Things I have tried:
Increasing the size of chessboard squares (48mm, 61mm, 109mm)
Increasing the number of rows/columns on the chessboard
Changing lighting conditions
More calibration images, I use a script to extract frames from a video so I normally use 20+ calibration frames
Using a smaller area to triangulate
Seeing if there was a change having one camera at an angle as opposed to having the cameras parallel
Checking the findChessboardCorner function actually finds the corners of the chessboard
Ensuring the chessboard is in many different positions (bottom corner, top corner, center of the frames for each camera)
Moving towards and away from the camera in videos.
Changed the criteria and criteria_stereo variables to see if that changed anything
Images of my most recent video, 109mm squares:
My code:
############### FIND CHESSBOARD CORNERS - OBJECT POINTS AND IMAGE POINTS #############################
chessboardSize = (8,5) # Other chessboard sizes used - (5,3) OR (9,6)
# Paths to the captured frames (should be in synch) (stereoLeft and stereoRight)
CALIBRATION_IMAGES_PATH_LEFT = 'images_vid\\stereoLeft\\*.png'
CALIBRATION_IMAGES_PATH_RIGHT = 'images_vid\\stereoRight\\*.png'
# termination criteria
criteria = (cv.TERM_CRITERIA_EPS + cv.TERM_CRITERIA_MAX_ITER, 30, 0.001)
objp = np.zeros((chessboardSize[0] * chessboardSize[1], 3), np.float32) # creates 9*6 list of (0.,0.,0.)
objp[:,:2] = np.mgrid[0:chessboardSize[0],0:chessboardSize[1]].T.reshape(-1,2) # formats list with (column no., row no., 0.) where max column no. = 8, and max row no. = 5
size_of_chessboard_squares_mm = 109
objp = objp * size_of_chessboard_squares_mm
# Arrays to store object points and image points from all the images.
objpoints = [] # 3d point in real world space
imgpointsL = [] # 2d points in image plane.
imgpointsR = [] # 2d points in image plane.
imagesLeft = sorted(glob.glob(CALIBRATION_IMAGES_PATH_LEFT))
imagesRight = sorted(glob.glob(CALIBRATION_IMAGES_PATH_RIGHT))
for imgLeft, imgRight in zip(imagesLeft, imagesRight):
imgL = cv.imread(imgLeft)
imgR = cv.imread(imgRight)
grayL = cv.cvtColor(imgL, cv.COLOR_BGR2GRAY)
grayR = cv.cvtColor(imgR, cv.COLOR_BGR2GRAY)
# Get the corners of the chess board
retL, cornersL = cv.findChessboardCorners(grayL, chessboardSize, None)
retR, cornersR = cv.findChessboardCorners(grayR, chessboardSize, None)
# Add object points and image points if chess board corners are found
if retL and retR == True:
objpoints.append(objp)
cornersL = cv.cornerSubPix(grayL, cornersL, (11,11), (-1,-1), criteria)
imgpointsL.append(cornersL)
cornersR = cv.cornerSubPix(grayR, cornersR, (11,11), (-1,-1), criteria)
imgpointsR.append(cornersR)
#Draw corners for user feedback
cv.drawChessboardCorners(imgL, chessboardSize, cornersL, retL)
cv.imshow('img left', imgL)
cv.drawChessboardCorners(imgR, chessboardSize, cornersR, retR)
cv.imshow('img right', imgR)
cv.waitKey()
cv.destroyAllWindows()
############# CALIBRATION #######################################################
retL, cameraMatrixL, distL, rvecsL, tvecsL = cv.calibrateCamera(objpoints, imgpointsL, frameSize, None, None)
heightL, widthL, channelsL = imgL.shape
newCameraMatrixL, roi_L = cv.getOptimalNewCameraMatrix(cameraMatrixL, distL, (widthL, heightL), 1, (widthL, heightL))
retR, cameraMatrixR, distR, rvecsR, tvecsR = cv.calibrateCamera(objpoints, imgpointsR, frameSize, None, None)
heightR, widthR, channelsR = imgR.shape
newCameraMatrixR, roi_R = cv.getOptimalNewCameraMatrix(cameraMatrixR, distR, (widthR, heightR), 1, (widthR, heightR))
######### Stereo Vision Calibration #############################################
## stereoCalibrate Output: retStereo is RSME, newCameraMatrixL and newCameraMatrixR are the intrinsic matrices for both
## cameras, distL and distR are the distortion coeffecients for both cameras, rot is the rotation matrix,
## trans is the translation matrix, and essentialMatrix and fundamentalMatrix are self descriptive
# R and T are taken from stereoCalibrate to use in triangulation
header = ['Rotation','Translation', 'ProjectionLeft', 'ProjectionRight'] # for the csv file
flags = 0
flags = cv.CALIB_FIX_INTRINSIC
criteria_stereo= (cv.TERM_CRITERIA_EPS + cv.TERM_CRITERIA_MAX_ITER, 100, 0.0001)
retStereo, newCameraMatrixL, distL, newCameraMatrixR, distR, rot, trans, essentialMatrix, fundamentalMatrix = cv.stereoCalibrate(objpoints, imgpointsL, imgpointsR, cameraMatrixL, distL, cameraMatrixR, distR, grayL.shape[::-1], criteria_stereo, flags)
print(retStereo)
print('stereo vision done')
Are there any immediate flags with my code or my calibration method? Or recommendations for improving the code? Thanks for taking your time to help me :)
In the image bellow, we see a defined world plane coordinate (X,Y,0) where Z=0. The camera as we can see is heading towards the defined world plane.
World reference point is located on the top left of the Grid (0,0,0). The distance between every two yellow point is 40 cm
I've calibrated my camera using the checkerboard and then used the built-in function cv2.solvePnP in order to estimate the rotation and translation vector of the camera with respect to my defined world coordinates. The results are as follows:
tvec_cam= [[-5.47884374]
[-3.08581371]
[24.15112048]]
rvec_cam= [[-0.02823308]
[ 0.08623225]
[ 0.01563199]]
According to the results, the (tx,ty,tz) seems to be right as the camera is located in the negative quarter of X,Y world-coordinates
However, i'm getting confused by interpreting the rotation vector.!
Does the resulted rotation vector say that the camera coordinates are almost aligned with the world coordinate axis (means almost no rotation!)?,
If yes how could this be true?, since according to OPENCV's camera coordinates, the Z-axis of the camera is pointing towards the scene (which means towards the world plane), the X-axis points towards the image write (which means opposite of X-world axis) and the Y-axis of the camera points towards the image bottom (which also means opposite of the Y-world axis)
Moreover, what is the unit of the tvec?
Note: I've illustrated the orientation of the defined world-coordinate axis according the the result of the translation vector (both tx and ty are negative)
the code i used for computing the rotation and translation vectors is shown below:
import cv2 as cv
import numpy as np
WPoints = np.zeros((9*3,3), np.float64)
WPoints[:,:2] = np.mgrid[0:9,0:3].T.reshape(-1,2)*0.4
imPoints=np.array([[20,143],[90,143],[161,143],[231,144],[303,144],
[374,144],[446,145],[516,146],[587,147],[18,214],[88,214]
,[159,215],[230,215],[302,216],[374,216],[446,216]
,[517,217],[588,217],[16,285],[87,285],[158,286],[229,287],
[301,288]
,[374,289],[446,289],[518,289],[589,289]],dtype=np.float64)
#load the rotation matrix [[4.38073915e+03 0.00000000e+00 1.00593352e+03]
# [0.00000000e+00 4.37829226e+03 6.97020491e+02]
# [0.00000000e+00 0.00000000e+00 1.00000000e+00]]
with np.load('parameters_cam1.npz') as X:
mtx, dist, _, _ = [X[i] for i in ('mtx','dist','rvecs','tvecs')]
ret,rvecs, tvecs = cv.solvePnP(WPoints, imPoints, mtx, dist)
np.savez("extrincic_camera1.npz",rvecs=rvecs,tvecs=tvecs)
print(tvecs)
print(rvecs)
cv.destroyAllWindows()
The code for estimating the intrinsic is show below
import numpy as np
import cv2
import glob
import argparse
import pathlib
ap = argparse.ArgumentParser()
ap.add_argument("-p", "--path", required=True, help="path to images folder")
ap.add_argument("-e", "--file_extension", required=False, default=".jpg",
help="extension of images")
args = vars(ap.parse_args())
path = args["path"] + "*" + args["file_extension"]
# termination criteria
criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 30, 0.001)
# prepare object points, like (0,0,0), (0.03,0,0), (0.06,0,0) ....,
#(0.18,0.12,0)
objp = np.zeros((5*7,3), np.float32)
objp[:,:2] = np.mgrid[0:7,0:5].T.reshape(-1,2)*0.03
#print(objp)
# Arrays to store object points and image points from all the images.
objpoints = [] # 3d point in real world space
imgpoints = [] # 2d points in image plane.
#images = glob.glob('left/*.jpg') #read a series of images
images = glob.glob(path)
path = 'foundContours'
#pathlib.Path(path).mkdir(parents=True, exist_ok=True)
found = 0
for fname in images:
img = cv2.imread(fname) # Capture frame-by-frame
#print(images[im_i])
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Find the chess board corners
ret, corners = cv2.findChessboardCorners(gray, (7,5), None)
# print(corners)
# If found, add object points, image points (after refining them)
if ret == True:
print('true')
objpoints.append(objp) # Certainly, every loop objp is the same, in 3D.
#print('obj_point',objpoints)
corners2 = cv2.cornerSubPix(gray,corners,(11,11),(-1,-1),criteria)
# print(corners2)
imgpoints.append(corners2)
print(imgpoints)
print('first_point',imgpoints[0])
#print(imgpoints.shape())
# Draw and display the corners
img = cv2.drawChessboardCorners(img, (7,5), corners2, ret)
found += 1
cv2.imshow('img', img)
cv2.waitKey(1000)
# if you want to save images with detected corners
# uncomment following 2 lines and lines 5, 18 and 19
image_name = path + '/calibresult' + str(found) + '.jpg'
cv2.imwrite(image_name, img)
print("Number of images used for calibration: ", found)
# When everything done, release the capture
# cap.release()
cv2.destroyAllWindows()
#calibration
ret, mtx, dist, rvecs, tvecs = cv2.calibrateCamera(objpoints, imgpoints,
gray.shape[::-1],None,None)
#save parameters needed in undistortion
np.savez("parameters_cam1.npz",mtx=mtx,dist=dist,rvecs=rvecs,tvecs=tvecs)
np.savez("points_cam1.npz",objpoints=objpoints,imgpoints=imgpoints)
print ("Camera Matrix = |fx 0 cx|")
print (" | 0 fy cy|")
print (" | 0 0 1|")
print (mtx)
print('distortion coefficients=\n', dist)
print('rotation vector for each image=', *rvecs, sep = "\n")
print('translation vector for each image=', *tvecs, sep= "\n")
Hope you could help me understanding this
Thanks in Advance
First, tvec is a in Axis-angle representation (https://en.wikipedia.org/wiki/Axis%E2%80%93angle_representation).
You can obtain the rotation matrix using cv2.Rodrigues(). For your data, I get almost the identity:
[[ 0.99616253 -0.01682635 0.08588995]
[ 0.01439347 0.99947963 0.02886672]
[-0.08633098 -0.02751969 0.99588635]]
Now, according to the directions of x and y in your picture, the z-axis points downwards (apply carefully the right-hand rule). This explains why the z-axis of the camera is almost aligned with the z-axis of your world reference frame.
Edit: Digging a little bit further, from the code you posted:
WPoints = np.zeros((9*3,3), np.float64)
WPoints[:,:2] = np.mgrid[0:9,0:3].T.reshape(-1,2)*0.4
The values for X and Y are all positive and increment to the right and to the bottom respectively, so you are indeed using the usual convention. You are actually using X and Y incrementing to the right and down respectively and what's wrong is only the arrows you drew in the picture.
Edit Concerning the interpretation of the translation vector: in the OpenCV convention, the points in the local camera reference frame are obtained from the points in the world reference frame like this:
|x_cam| |x_world|
|y_cam| = Rmat * |y_world| + tvec
|z_cam| |z_world|
With this convention, tvec is the position of the world origin in the camera reference frame. What's more easily interpretable is the position of the camera origin in the world reference frame, which can be obtained as:
cam_center = -(tvec * R_inv)
Where R_inv is the inverse of the rotation matrix. Here the rotation matrix is almost the identity, so a quick approximation would be -tvec, which is (5.4, 3.1, -24.1).
I'm working on depth map with OpenCV. I can obtain it but it is reconstructed from the left camera origin and there is a little tilt of this latter and as you can see on the figure, the depth is "shifted" (the depth should be close and no horizontal gradient):
I would like to express it as with a zero angle, i try with the warp perspective function as you can see below but i obtain a null field...
P = np.dot(cam,np.dot(Transl,np.dot(Rot,A1)))
dst = cv2.warpPerspective(depth, P, (2048, 2048))
with :
#Projection 2D -> 3D matrix
A1 = np.zeros((4,3))
A1[0,0] = 1
A1[0,2] = -1024
A1[1,1] = 1
A1[1,2] = -1024
A1[3,2] = 1
#Rotation matrice around the Y axis
theta = np.deg2rad(5)
Rot = np.zeros((4,4))
Rot[0,0] = np.cos(theta)
Rot[0,2] = -np.sin(theta)
Rot[1,1] = 1
Rot[2,0] = np.sin(theta)
Rot[2,2] = np.cos(theta)
Rot[3,3] = 1
#Translation matrix on the X axis
dist = 0
Transl = np.zeros((4,4))
Transl[0,0] = 1
Transl[0,2] = dist
Transl[1,1] = 1
Transl[2,2] = 1
Transl[3,3] = 1
#Camera Intrisecs matrix 3D -> 2D
cam = np.concatenate((C1,np.zeros((3,1))),axis=1)
cam[2,2] = 1
P = np.dot(cam,np.dot(Transl,np.dot(Rot,A1)))
dst = cv2.warpPerspective(Z0_0, P, (2048*3, 2048*3))
EDIT LATER :
You can download the 32MB field dataset here: https://filex.ec-lille.fr/get?k=cCBoyoV4tbmkzSV5bi6. Then, load and view the image with:
from matplotlib import pyplot as plt
import numpy as np
img = np.load('testZ0.npy')
plt.imshow(img)
plt.show()
I have got a rough solution in place. You can modify it later.
I used the mouse handling operations available in OpenCV to crop the region of interest in the given heatmap.
(Did I just say I used a mouse to crop the region?) Yes, I did. To learn more about mouse functions in OpenCV SEE THIS. Besides, there are many other SO questions that can help you in this regard.:)
Using those functions I was able to obtain the following:
Now to your question of removing the tilt. I used the homography principal by taking the corner points of the image above and using it on a 'white' image of a definite size. I used the cv2.findHomography() function for this.
Now using the cv2.warpPerspective() function in OpenCV, I was able to obtain the following:
Now you can the required scale to this image as you wanted.
CODE:
I have also attached some snippets of code for your perusal:
#First I created an image of white color of a definite size
back = np.ones((435, 379, 3)) # size
back[:] = (255, 255, 255) # white color
Next I obtained the corner points pts_src on the tilted image below :
pts_src = np.array([[25.0, 2.0],[403.0,22.0],[375.0,436.0],[6.0,433.0]])
I wanted the points above to be mapped to the points 'pts_dst' given below :
pts_dst = np.array([[2.0, 2.0], [379.0, 2.0], [379.0, 435.0],[2.0, 435.0]])
Now I used the principal of homography:
h, status = cv2.findHomography(pts_src, pts_dst)
Finally I mapped the original image to the white image using perspective transform.
fin = cv2.warpPerspective(img, h, (back.shape[1],back.shape[0]))
# img -> original tilted image.
# back -> image of white color.
Hope this helps! I also got to learn a great deal from this question.
Note: The points fed to the 'cv2.findHomography()' must be in float.
For more info on Homography , visit THIS PAGE
I am trying to simulate an image standing out of a marker. This is my code so far which does what is pictured. Essentially, I just want to rotate the image to appear to stand out orthogonal to the checkerboard.
As you can see, I use the code to find the transformation matrix between a normalized square image and the corresponding checkerboard corners. I then use warpPerspective to get the image you see. I know that I can use the rotation vectors from the solvePnP to obtain a rotation matrix through rodrigues() but I dont know what the next step is
def transformTheSurface(inputFrame):
ret, frameLeft = capleft.read()
capGray = cv2.cvtColor(frameLeft,cv2.COLOR_BGR2GRAY)
found, corners = cv2.findChessboardCorners(capGray, (5,4), None, cv2.CALIB_CB_NORMALIZE_IMAGE + cv2.CALIB_CB_ADAPTIVE_THRESH ) #,None,cv2.CALIB_CB_FAST_CHECK)
if (found):
npGameFrame = pygame.surfarray.array3d(inputFrame)
inputFrameGray = cv2.cvtColor(npGameFrame,cv2.COLOR_BGR2GRAY)
cv2.drawChessboardCorners(frameLeft, (5,4), corners, found)
q = corners[[0, 4, 15, 19]]
ret, rvecs, tvecs = cv2.solvePnP(objp, corners, mtx, dist)
ptMatrix = cv2.getPerspectiveTransform( muffinCoords, q)
npGameFrame = cv2.flip(npGameFrame, 0)
ptMatrixWithXRot = ptMatrix * rodRotMat[0]
#inputFrameConv = cv2.cvtColor(npGameFrame,cv2.COLOR_BGRA2GRAY)
transMuffin = cv2.warpPerspective(npGameFrame, ptMatrix, (640, 480)) #, muffinImg, cv2.INTER_NEAREST, cv2.BORDER_CONSTANT, 0)
EDIT:
I have added some more code, in hopes to create my own 3x3 transformation matrix. I used the following reference . Here is my code for that:
#initialization happens earlier in code
muffinCoords = np.zeros((4,2), np.float32)
muffinCoords[0] = (0,0)
muffinCoords[1] = (200,0)
muffinCoords[2] = (0,200)
muffinCoords[3] = (200,200)
A1 = np.zeros((4,3), np.float32)
A1[0] = (1,0,322)
A1[1] = (0,1,203)
A1[2] = (0,0,0)
A1[3] = (0,0,1)
R = np.zeros((4,4), np.float32)
R[3,3] = 1.0
T = np.zeros((4,4), np.float32)
T[0] = (1,0,0,0)
T[1] = (0,1,0,0)
T[2] = (0,0,1,0)
T[3] = (0,0,0,1)
#end initialization
#load calib data derived using cv2.calibrateCamera, my Fx and Fy are about 800
loadedCalibFileMTX = np.load('calibDataMTX.npy')
mtx = np.zeros((3,4), np.float32)
mtx[:3,:3] = loadedCalibFileMTX
#this is new to my code, creating what I interpret as Rx*Ry*Rz
ret, rvecCalc, tvecs = cv2.solvePnP(objp, corners, loadedCalibFileMTX, dist)
rodRotMat = cv2.Rodrigues(rvecCalc)
R[:3,:3] = rodRotMat[0]
#then I create T
T[0,3] = tvecs[0]
T[1,3] = tvecs[1]
T[2,3] = tvecs[2]
# CREATING CUSTOM TRANSFORM MATRIX
# A1 -> 2d to 3d projection matrix
# R-> rotation matrix as calculated by solve PnP, or Rx * Ry * Rz
# T -> converted translation matrix, reference from site, vectors pulled from tvecs of solvPnP
# mtx -> 3d to 2d matrix
# customTransformMat = mtx * (T * (R * A1)) {this is intended calculation of following}
first = np.dot(R, A1)
second = np.dot(T, first)
finalCalc = np.dot(mtx, second)
finalNorm = finalCalc/(finalCalc[2,2]) # to make sure that the [2,2] element is 1
transMuffin = cv2.warpPerspective(npGameFrame, finalNorm, (640, 480), None, cv2.INTER_NEAREST, cv2.BORDER_CONSTANT, 0)
#transMuffin is returned as undefined here, any help?
# using the cv2.getPerspectiveTransform method to find what you can find pictured at the top
ptMatrix = cv2.getPerspectiveTransform( muffinCoords, q)
I finally figured out the right methodology. you can find the code here https://github.com/mikezucc/augmented-reality-fighter-pygame
NOTICE:
Almost ALL of the game code is written by Leif Theiden, and is under license specified in the .py files. The code relevant to the computer vision is in states.py. I used the game to just show that it can be done for others looking to get started in simple computer vision.
My code opens a thread everytime a new surface (PyGame for frame simply) is called to be displayed on the main window. I start a thread at that point and execute a simple computer vision function that does the following:
Searches a camera stream frame for the 5x4 chessboard (cv2.findChessboardCorners)
The found corners are then drawn onto the image
Using cv2.solvePnP, the approximate pose (Rotation and translation vectors) are derived
The 3d points that describe a square are then projected from the 3d space determined by step 3 into a 2d space. this is used to convert a predertimined 3d structure into something you can use to graph on a 2d image.
However, this step instead finds the transformation to get from a set of 2d square points (the dimensions of the game frame) to the newly found projected 2d points (of the 3d frame). Now you can see that what we are trying to is simply do a two step transformation.
I then perform a basic tutorial style addition of the captured stream frame and the transformed game frame to get a final image
Variables:
+from3dTransMatrix -> points of the projected 3d structure into 2d points. these are the red dots you see
+q -> this is the reference plane that we determine the pose from
+ptMatrix -> the final transformation, to transform the game frame to fit in the projected frame
Check out the screens in the topmost folder ;]
enjoy!
I'm trying to look for shapes in an image using OpenCV. I know the shapes I want to match (there are some shapes I don't know about, but I don't need to find them) and their orientations. I don't know their sizes (scale) and locations.
My current approach:
Detect contours
For each contour, calculate the maximum bounding box
Match each bounding box to one of the known shapes separately. In my real project, I'm scaling the region to the template size and calculating differences in Sobel gradient, but for this demo, I'm just using the aspect ratio.
Where this approach comes undone is where shapes touch. The contour detection picks up the two adjacent shapes as a single contour (single bounding box). The matching step will then obviously fail.
Is there a way to modify my approach to handle adjacent shapes separately? Also, is there a better way to perform step 3?
For example: (Es colored green, Ys colored blue)
Failed case: (unknown shape in red)
Source code:
import cv
import sys
E = cv.LoadImage('e.png')
E_ratio = float(E.width)/E.height
Y = cv.LoadImage('y.png')
Y_ratio = float(Y.width)/Y.height
EPSILON = 0.1
im = cv.LoadImage(sys.argv[1], cv.CV_LOAD_IMAGE_GRAYSCALE)
storage = cv.CreateMemStorage(0)
seq = cv.FindContours(im, storage, cv.CV_RETR_EXTERNAL,
cv.CV_CHAIN_APPROX_SIMPLE)
regions = []
while seq:
pts = [ pt for pt in seq ]
x, y = zip(*pts)
min_x, min_y = min(x), min(y)
width, height = max(x) - min_x + 1, max(y) - min_y + 1
regions.append((min_x, min_y, width, height))
seq = seq.h_next()
rgb = cv.LoadImage(sys.argv[1], cv.CV_LOAD_IMAGE_COLOR)
for x,y,width,height in regions:
pt1 = x,y
pt2 = x+width,y+height
if abs(float(width)/height - E_ratio) < EPSILON:
color = (0,255,0,0)
elif abs(float(width)/height - Y_ratio) < EPSILON:
color = (255,0,0,0)
else:
color = (0,0,255,0)
cv.Rectangle(rgb, pt1, pt2, color, 2)
cv.ShowImage('rgb', rgb)
cv.WaitKey(0)
e.png:
y.png:
good:
bad:
Before anybody asks, no, I'm not trying to break a captcha :) OCR per se isn't really relevant here: the actual shapes in my real project aren't characters -- I'm just lazy, and characters are the easiest thing to draw (and still get detected by trivial methods).
As your shapes can vary in size and ratio, you should look at scaling invariant descriptors. A bunch of such descriptors would be perfect for your application.
Process those descriptors on your test template and then use some kind of simple classification to extract them. It should give pretty good results with simple shapes as you show.
I used Zernike and Hu moments in the past, the latter being the most famous. You can find an example of implementation here : http://www.lengrand.fr/2011/11/classification-hu-and-zernike-moments-matlab/.
Another thing : Given your problem, you should look at OCR technologies (stands for optical character recognition : http://en.wikipedia.org/wiki/Optical_character_recognition ;)).
Hope this helps a bit.
Julien
Have you try Chamfer Matching or contour matching (correspondence) using CCH as descriptor.
Chamfer matching is using distance transform of target image and template contour. not exactly scale invariant but fast.
The latter is rather slow, as the complexity is at least quadratic for bipartite matching problem. on the other hand, this method is invariant to scale, rotation, and probably local distortion (for approximate matching, which IMHO is good for the bad example above).