Ho provato a vedere se effettuare la calibrazione di tag Aruco con immagini in formato raw lossless od in formato jpg (lossy) poteva influenzare l'errore finale
Per fare cio' ho preso un telefono Android che permette di salvare la stessa immagine in formato DNG ed in formato JPG
Lo script di calcolo e' il seguente
import numpy as np
import cv2
import cv2.aruco as aruco
# Aruco parameters
aruco_dict = aruco.Dictionary_get(aruco.DIC
aruco_params = aruco.DetectorParameters_creat
# Chessboard parameters
num_chessboard_corners_horizon
num_chessboard_corners_vertica
chessboard_square_size = 0.025 # in meters
# Image directory and file extension
image_directory = 'path_to_images_directory/'
image_extension = '.jpg'
# Arrays to store object points and image points
obj_points = [] # 3D points in real-world coordinates
img_points = [] # 2D points in image plane
# Generate chessboard object points
chessboard_size = (num_chessboard_corners_horizo
num_chessboard_corners_vertica
objp = np.zeros((num_chessboard_corne
num_chessboard_corners_vertica
objp[:, :2] = np.mgrid[0:num_chessboard_corn
0:num_chessboard_corners_verti
objp = objp * chessboard_square_size
# Iterate through images and find chessboard corners
image_count = 0
while True:
# Load image
image_path = image_directory + str(image_count) + image_extension
image = cv2.imread(image_path)
if image is None:
break
# Convert to grayscale
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# Find chessboard corners
ret, corners = cv2.findChessboardCorners(gray
# If corners found, add object points and image points
if ret == True:
obj_points.append(objp)
img_points.append(corners)
# Draw and display corners
cv2.drawChessboardCorners(imag
cv2.imshow('Chessboard Corners', image)
cv2.waitKey(500)
image_count += 1
cv2.destroyAllWindows()
# Perform Aruco calibration
ret, mtx, dist, rvecs, tvecs = cv2.calibrateCamera(obj_points
img_points, gray.shape[::-1], None, None)
# Calculate reprojection error
mean_error = 0
for i in range(len(obj_points)):
img_points_proj, _ = cv2.projectPoints(obj_points[i
tvecs[i], mtx, dist)
error = cv2.norm(img_points[i], img_points_proj, cv2.NORM_L2) /
len(img_points_proj)
mean_error += error
calibration_error = mean_error / len(obj_points)
print("Calibration error: {}".format(calibration_error))
# Estimate additional error
reproj_error = 0
total_points = 0
for i in range(len(obj_points)):
img_points_proj, _ = cv2.projectPoints(obj_points[i
tvecs[i], mtx, dist)
error = cv2.norm(img_points[i], img_points_proj, cv2.NORM_L2)
reproj_error += error
total_points += len(obj_points[i])
additional_error = reproj_error / total_points
print("Additional error: {}".format(additional_error))
I risultati sono i seguenti (per i file DNG ho usato la libreria rawpy)
JPG
python calibration.py --dir ./jpg/ --square_size 0.015 -w 9 -t 6
Calibration error: 0.13602731796497697
Additional error: 0.13602731796497697
DNG
Calibration error: 0.13606447706803765
Additional error: 0.13606447706803765
In conclusione non si ha una significativa differenza di errore nell'usare JPG o DNG