def build_trainset(input_dir="data", output_dir="trainset", win_size=321):
'''scan the input folder and put every image with annotations
output folder'''
utils.init_face_detector(True, win_size)
utils.load_shape_predictor("dlib/shape_predictor_68_face_landmarks.dat")
count = int(utils.count_files_inside(output_dir) / 8)
window = "window"
cv2.namedWindow(window)
for face, box in utils.faces_inside(input_dir, 1, True):
face_copy = face.copy()
face_rect = region(face)
# detections
points = utils.detect_landmarks(face, face_rect)
utils.draw_points(face, points)
# show face
while (1):
h, w = face.shape[:2]
if h > 0 and w > 0:
cv2.imshow(window, show_properly(face))
else:
break
key = cv2.waitKey(20) & 0Xff
if key == Keys.ESC:
break # skip current face
elif key == Keys.S:
path = f"{output_dir}/face{count}"
# save image
cv2.imwrite(f"{path}.jpg", face_copy)
# save annotation relative to the current face
Annotation(f"{path}.ann", face_rect, points).save()
# generate and save augumetations
array = augment_data(face_copy, face_rect, points)
for i, x in enumerate(array):
# save image x[0]
cv2.imwrite(f"{path}_{i + 1}.jpg", x[0])
# save annotations
Annotation(f"{path}_{i + 1}.ann", face_rect, x[1]).save()
count = count + 1
break
elif key == Keys.Q:
# quit program
return cv2.destroyAllWindows()
cv2.destroyAllWindows()
def compare_models(folder="testset", m1=None, m2=None, view=False):
'''compare the [m1] shape_predictor aganist the [m2] model,
optionally you can [view] the results'''
assert (m1 and m2)
utils.init_face_detector(True, 150)
# load models
utils.load_shape_predictor(m2)
sp_m2 = utils.shape_predictor
utils.load_shape_predictor(m1)
sp_m1 = utils.shape_predictor
# init error
err = 0
num = 0
for face, region in utils.faces_inside(folder):
h, w = face.shape[:2]
if h == 0 or w == 0:
continue
box = utils.Region(0, 0, region.width, region.height)
# detect landmarks
utils.shape_predictor = sp_m1
lmarks_m1 = utils.detect_landmarks(face, box)
utils.shape_predictor = sp_m2
lmarks_m2 = utils.detect_landmarks(face, box)
# update error:
num += 1
# err += normalized_root_mean_square(lmarks_m1, lmarks_m2)
# results:
if view is True:
utils.draw_points(face, lmarks_m1, color=Colors.green)
utils.draw_points(face, lmarks_m2, color=Colors.red)
utils.show_image(utils.show_properly(face))
if num != 0:
err /= num
print("the NRMSE of m1 aganist m2 is {}".format(err))
def test(folder="testset", model="dlib/shape_predictor_68_face_landmarks.dat"):
utils.init_face_detector(True, 150)
utils.load_shape_predictor(model)
my_sp = utils.shape_predictor
dlib_sp = dlib.shape_predictor(
"dlib/shape_predictor_68_face_landmarks.dat")
for face, r in utils.faces_inside(folder):
box = region(face)
utils.shape_predictor = my_sp
lmarks0 = utils.detect_landmarks(face, box)
utils.shape_predictor = dlib_spq
lmarks1 = utils.detect_landmarks(face, box)
# draw results
utils.draw_points(face, lmarks1, color=Colors.green)
utils.draw_points(face, lmarks0, color=Colors.red)
utils.show_image(show_properly(face))
def test_augment():
utils.init_face_detector(True, 321)
utils.load_shape_predictor("dlib/shape_predictor_68_face_landmarks.dat")
for img in utils.images_inside("trainset"):
points = utils.detect_landmarks(img, region(img))
angle = 30
h, w = img.shape[:2]
center = (w / 2, h / 2)
# 30 degree rotation
rot1 = utils.rotate_image(img, angle)
rot_pts1 = utils.rotate_landmarks(points, center, angle)
# -30 degree rotatation
rot2 = utils.rotate_image(img, -angle)
rot_pts2 = utils.rotate_landmarks(points, center, -angle)
# mirroring
mir = utils.flip_image(img)
mir_pts = utils.detect_landmarks(mir, region(mir))
utils.draw_points(img, points)
utils.draw_points(rot1, rot_pts1, color=Colors.cyan)
utils.draw_points(rot2, rot_pts2, color=Colors.purple)
utils.draw_points(mir, mir_pts, color=Colors.green)
while True:
cv2.imshow("image", img)
cv2.imshow("mirrored", mir)
cv2.imshow("rotated30", rot1)
cv2.imshow("rotated-30", rot2)
key = cv2.waitKey(20) & 0Xff
if key == Keys.ESC:
break
elif key == Keys.Q:
return cv2.destroyAllWindows()
def main():
global image, boxes, points
# getting arguments from cli
args = utils.cli_arguments()
out = utils.open_file(args)
img_dir = args["dir"]
auto_landm = args["land"] is not None
auto_faces = args["auto"]
mirror_points = args["mirror"]
if args["train"] is True:
# skip everything and train the model
return utils.train_model(out.path, args["train"])
# cnn face detector
utils.init_face_detector(args)
# load shape predictor if requested
if auto_faces and auto_landm:
global predictor
predictor = dlib.shape_predictor(args["land"])
# recover last state (if append is true)
resume, lastpath = utils.load_state(args["append"])
for file in os.listdir(img_dir):
# consider only images
if not is_image(file):
continue
# avoid mirrored images
if is_mirrored(file):
continue
# load image:
path = os.path.join(img_dir, file)
# trying to resume from the image located at lastpath
if resume:
if path == lastpath:
resume = False
else:
continue
image = cv2.imread(path)
# clear: stack, boxes and points
stack.clear()
boxes.clear()
points.clear()
# automatically detect faces
if auto_faces:
draw_face(image)
# and landmarks
if auto_landm:
detect_landmarks(image, boxes[-1])
stack.append(image)
# create a window with disabled menu when right-clicking with the mouse
window = file
cv2.namedWindow(window, cv2.WINDOW_AUTOSIZE | cv2.WINDOW_GUI_NORMAL)
# mouse callback to window
cv2.setMouseCallback(window, mouse_callback)
# removing or skipping the current image without affecting output file
skipped = False
removed = False
# showing image until esc is pressed
while (1):
cv2.imshow(window, image)
key = cv2.waitKey(20) & 0Xff
# listen to key events
if key == KEY_ESC:
break
elif key == KEY_S:
skipped = True
break
elif key == KEY_R:
removed = True
utils.delete_image(path)
break
elif key == KEY_Q:
return quit(path, out)
if not (skipped or removed):
# clear point log
print()
# write annotations
add_entry(out, path, boxes, points, mirror_points)
# close window
cv2.destroyAllWindows()
# delete checkpoint file and close output file
utils.delete_state()
out.close()
def build_trainset_auto(src="dataset", dst="trainset", debug=False):
'''build a trainset automatically from an ibug-like dataset,
the images are taken from [src] folder and saved to [dst] folder'''
utils.init_face_detector(True, 150)
qualiy = [int(cv2.IMWRITE_JPEG_QUALITY), 50]
# file count for naming
count = int(utils.count_files_inside(dst) / 8)
for img, lmarks, path in utils.ibug_dataset(src):
h, w = img.shape[:2]
# crop a bigger region around landmarks
region = utils.points_region(lmarks)
scaled = region.scale(1.8, 1.8).ensure(w, h)
img = utils.crop_image(img, scaled)
# detect faces
face = utils.prominent_face(utils.detect_faces(img))
# if cnn fails try with dlib
if face is None:
faces = utils.detect_faces(img, detector="dlib")
# ..if dlib fails take the region around landmarks
if face is None:
face = region.copy()
else:
face = utils.prominent_face(faces)
# edit landmarks according to scaled region
lmarks = adjust_landmarks(scaled, lmarks)
# augumentations
i = 0
for image, landmarks, box in augment_data(img, face, lmarks):
i = i + 1
if debug:
utils.draw_rect(image, box, color=Colors.yellow)
utils.draw_points(image, landmarks, color=Colors.purple)
name = f"image{i}"
utils.show_image(show_properly(image), window=name)
else:
# save annotation and image
ipath = os.path.join(dst, f"face{count}_{i}.jpg")
apath = os.path.join(dst, f"face{count}_{i}.ann")
cv2.imwrite(ipath, image, qualiy)
Annotation(apath, box.as_list()[:4], landmarks).save()
if debug:
utils.draw_rect(img, face, color=Colors.red)
utils.draw_points(img, lmarks, color=Colors.green)
utils.show_image(show_properly(img))
else:
# save image and annotation
ipath = os.path.join(dst, f"face{count}.jpg")
apath = os.path.join(dst, f"face{count}.ann")
cv2.imwrite(ipath, img, qualiy)
Annotation(apath, face.as_list()[:4], lmarks).save()
count = count + 1
# info
print("{} processed: {}\r".format(count, ipath))