def compute_rotation(contour1, contour2):
resolution = 50
scale = 0.5
val = 1 # set to 1 for use, 255 for debugging
shape1 = contour_to_normalized_points(contour1, resolution, scale)
shape2 = contour_to_normalized_points(contour2, resolution, scale)
# print shape1
num_angles = 15
img1 = np.zeros((resolution, resolution), np.uint8)
cv2.fillPoly(img1, [shape1], val)
img1_flipped = cv2.flip(img1, 1)
show(img1)
similarities = []
transforms = []
for rotation in [x * math.pi * 2.0 / num_angles for x in xrange(num_angles)]:
for flip in (False, True):
transforms.append((rotation, flip))
img2 = np.zeros((resolution, resolution), np.uint8)
cv2.fillPoly(img2, [shape2], val)
for rotation, flip in transforms:
M = cv2.getRotationMatrix2D((resolution / 2, resolution / 2), rotation * 180.0 / math.pi, 1.0)
img2_rotated = cv2.warpAffine(img2, M, (resolution, resolution))
show(img2_rotated)
overlap = np.multiply((img1_flipped if flip else img1), img2_rotated)
x = np.sum(overlap) * 1.0 / (resolution ** 2)
similarities.append(x)
best_idx = max(range(len(similarities)), key=lambda i: similarities[i])
best_non_flipped_idx = max(range(0, len(similarities), 2), key=lambda i: similarities[i])
rot, flipped = transforms[best_idx]
rot_without_flipping = transforms[best_non_flipped_idx][0]
return rot, flipped, rot_without_flipping
def photo_demo():
# Capture frame-by-frame
frame = cv2.imread("images/camera-view.png")
frame = cv2.resize(frame, IMAGE_SIZE)
show(frame, 'image')
# frame = cv2.undistort(frame, mtx, dist, None, newcameramtx)
frame = cv2.warpPerspective(frame, perspective_matrix, IMAGE_SIZE)
frame = cv2.flip(frame, 1)
frame = cv2.flip(frame, -1)
show(frame, 'transformed')
frame, output = process_image(frame)
global IMAGE
IMAGE = frame
global OUTPUT
OUTPUT = output
if SHOW_SINGLE_CAPTURE_AND_DUMP:
cv2.imshow('label', frame)
cv2.waitKey(0)
# cap.release()
cv2.destroyAllWindows()
def photo_demo():
# Capture frame-by-frame
frame = cv2.imread("images/camera-view.png")
frame = cv2.resize(frame, IMAGE_SIZE)
show(frame, 'image')
# frame = cv2.undistort(frame, mtx, dist, None, newcameramtx)
frame = cv2.warpPerspective(frame, perspective_matrix, IMAGE_SIZE)
frame = cv2.flip(frame, 1)
frame = cv2.flip(frame, -1)
show(frame, 'transformed')
frame, output = process_image(frame)
global IMAGE
IMAGE = frame
global OUTPUT
OUTPUT = output
if SHOW_SINGLE_CAPTURE_AND_DUMP:
cv2.imshow('label', frame)
cv2.waitKey(0)
# cap.release()
cv2.destroyAllWindows()
def video_loop():
cap = cv2.VideoCapture(CAMERA_NUM)
while(True):
# Capture frame-by-frame
if BLANKING:
blanking.set_blank(True)
wait(0.1)
if BLANKING:
blanking.set_blank(False)
wait(0.1)
cap.grab()
# blanking.set_blank(False)
ret, frame = cap.retrieve()
frame = cv2.resize(frame, IMAGE_SIZE)
# frame = cv2.undistort(frame, mtx, dist, None, newcameramtx)
frame = cv2.warpPerspective(frame, perspective_matrix, IMAGE_SIZE)
frame = cv2.flip(frame, 1)
frame = cv2.flip(frame, -1)
frame, output = process_image(frame)
cv2.imshow('image', frame)
global IMAGE
IMAGE = frame
global OUTPUT
OUTPUT = output
if SHOW_SINGLE_CAPTURE_AND_DUMP:
cv2.waitKey(0)
break
if BLANKING:
wait(0.7)
else:
wait(0.01)
cap.release()
cv2.destroyAllWindows()
def compute_rotation(contour1, contour2):
resolution = 50
scale = 0.5
val = 1 # set to 1 for use, 255 for debugging
shape1 = contour_to_normalized_points(contour1, resolution, scale)
shape2 = contour_to_normalized_points(contour2, resolution, scale)
# print shape1
num_angles = 15
img1 = np.zeros((resolution, resolution), np.uint8)
cv2.fillPoly(img1, [shape1], val)
img1_flipped = cv2.flip(img1, 1)
show(img1)
similarities = []
transforms = []
for rotation in [
x * math.pi * 2.0 / num_angles for x in xrange(num_angles)
]:
for flip in (False, True):
transforms.append((rotation, flip))
img2 = np.zeros((resolution, resolution), np.uint8)
cv2.fillPoly(img2, [shape2], val)
for rotation, flip in transforms:
M = cv2.getRotationMatrix2D((resolution / 2, resolution / 2),
rotation * 180.0 / math.pi, 1.0)
img2_rotated = cv2.warpAffine(img2, M, (resolution, resolution))
show(img2_rotated)
overlap = np.multiply((img1_flipped if flip else img1), img2_rotated)
x = np.sum(overlap) * 1.0 / (resolution**2)
similarities.append(x)
best_idx = max(range(len(similarities)), key=lambda i: similarities[i])
best_non_flipped_idx = max(range(0, len(similarities), 2),
key=lambda i: similarities[i])
rot, flipped = transforms[best_idx]
rot_without_flipping = transforms[best_non_flipped_idx][0]
return rot, flipped, rot_without_flipping
