def calculate_similarity(self, image1, image2):
bboxes1 = self.face_detector.detect(image1)
bboxes2 = self.face_detector.detect(image2)
facial5points1 = [[bboxes1[0][2 * i - 1], bboxes1[0][2 * i]] for i in range(3, 8)]
warped_face1 = warp_and_crop_face(
cv2.cvtColor(image1, cv2.COLOR_BGR2RGB), facial5points1, self.reference, (self.crop_size, self.crop_size))
facial5points2 = [[bboxes2[0][2 * i - 1], bboxes2[0][2 * i]] for i in range(3, 8)]
warped_face2 = warp_and_crop_face(
cv2.cvtColor(image2, cv2.COLOR_BGR2RGB), facial5points2, self.reference, (self.crop_size, self.crop_size))
input1 = self.preprocess(warped_face1)
input2 = self.preprocess(warped_face2)
embedding1 = self.arcface_r50_asian(input1)
embedding2 = self.arcface_r50_asian(input2)
embedding1 = embedding1.detach().numpy()
print(embedding1.shape)
embedding2 = embedding2.detach().numpy()
cosin = cosine_similarity(self.l2_normalize(embedding1), self.l2_normalize(embedding2))
distance = euclidean_distances(embedding1, embedding2)
return cosin[0, 0], distance[0, 0], warped_face1, warped_face2
def warpface(self, landmarks, img):
faces = []
for landmark in landmarks:
facial5points = [[landmark[j], landmark[j + 5]] for j in range(5)]
face = warp_and_crop_face(np.array(img), facial5points, self.reference, crop_size=(self.crop_size, self.crop_size))
faces.append(torchvision.transforms.ToTensor()(face))
return torch.stack(faces)
def detect_one_face_align(img):
crop_size = 112
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
#img = Image.open(os.path.join(image_path))
landmarks = []
bounding_boxes = []
try:
bounding_boxes, landmarks = detect_faces(img)
except Exception as e:
print(e)
if len(
landmarks
) == 0: # If the landmarks cannot be detected, the img will be discarded
return None
#print(landmarks)
facial5point = [[landmarks[0][j], landmarks[0][j + 5]] for j in range(5)]
#print(facial5points[i])
warped_face = warp_and_crop_face(np.array(img),
facial5point,
reference,
crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
# img_warped.save("test.jpg")
# img_warped.show()
return img_warped, bounding_boxes
def process_faces(img, det_models, reference, crop_size):
"""
Note that img is a PIL Image!
"""
try: # Handle exception
bounding_boxes, landmarks = detect_faces(det_models, img)
except Exception as e:
print("Image is discarded due to exception:\n{}".format(e))
return []
# If the landmarks cannot be detected, the img will be discarded
if len(landmarks) == 0:
print("No landmarks detected!")
return []
else:
warped_faces = []
for i, landmark in enumerate(landmarks):
facial5points = [[landmark[j], landmark[j + 5]] for j in range(5)]
warped_faces.append(
warp_and_crop_face(np.array(img),
facial5points,
reference,
crop_size=(crop_size, crop_size)))
assert len(warped_faces) == len(bounding_boxes)
assert len(warped_faces) == len(landmarks)
return [
FaceResult(
bounding_box=bounding_boxes[i],
landmark=landmarks[i],
warped_face=warped_faces[i],
) for i in range(len(warped_faces))
]
def detect_face_align(img):
crop_size = 112 #因为backbone的input_size要[112,112]or[224,224]
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
#img = Image.open(os.path.join(image_path))
bounding_boxes = []
warped_face = []
img_warped = []
facial5point = []
landmarks = []
try:
bounding_boxes, landmarks = detect_faces(img)
except Exception as e:
print(e)
if len(
landmarks
) == 0: # If the landmarks cannot be detected, the img will be discarded
return img_warped, bounding_boxes
#print(landmarks)
for i in range(len(landmarks)):
facial5point.append([[landmarks[i][j], landmarks[i][j + 5]]
for j in range(5)])
#print(facial5points[i])
warped_face.append(
warp_and_crop_face(np.array(img),
facial5point[i],
reference,
crop_size=(crop_size, crop_size)))
img_warped.append(Image.fromarray(warped_face[i]))
# img_warped.save("test.jpg")
# img_warped.show()
return img_warped, bounding_boxes
def align_68(img, crop_size, model):
# check if img is PIL
bbox = model.face_detector.detect_from_image(img)
if bbox is None:
return 4
landmarks = model.get_landmarks(img, bbox)
if landmarks is None:
return 4
landmarks = landmarks[0]
left_eye = np.mean(landmarks[36:42, :], axis=0).astype(np.int)
right_eye = np.mean(landmarks[42:48, :], axis=0).astype(np.int)
nose = np.mean(landmarks[28:35, :], axis=0).astype(np.int)
left_mouth = landmarks[48].astype(np.int)
right_mouth = landmarks[54].astype(np.int)
# settings
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
facial5points = [left_eye, right_eye, nose, left_mouth, right_mouth]
warped_face = warp_and_crop_face(np.array(img),
facial5points,
reference,
crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
return img_warped
def align_face(img, landmarks, crop_size=112):
"""Align face on the photo
Arguments:
img {PIL.Image} -- Image with face
landmarks {np.array} -- Key points
Keyword Arguments:
crop_size {int} -- Size of face (default: {112})
Returns:
PIL.Image -- Aligned face
"""
facial5points = [[landmarks[j], landmarks[j + 5]] for j in range(5)]
warped_face = warp_and_crop_face(np.array(img), facial5points, reference, crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
return img_warped
def align(img, crop_size, img_landmarks):
landmarks = img_landmarks[0]
left_eye = np.mean(landmarks[36:42, :], axis=0).astype(np.int)
right_eye = np.mean(landmarks[42:48, :], axis=0).astype(np.int)
nose = np.mean(landmarks[28:35, :], axis=0).astype(np.int)
left_mouth = landmarks[48].astype(np.int)
right_mouth = landmarks[54].astype(np.int)
# settings
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
facial5points = [left_eye, right_eye, nose, left_mouth, right_mouth]
warped_face = warp_and_crop_face(np.array(img),
facial5points,
reference,
crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
return img_warped
def align_(img, crop_size, model_root, filter=False):
# check if img is PIL
if type(img) == np.ndarray:
img = Image.fromarray(img)
# path to detector models
op = "{}/onet.npy".format(model_root)
pp = "{}/pnet.npy".format(model_root)
rp = "{}/rnet.npy".format(model_root)
# settings
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
try: # Handle exception
_, landmarks = detect_faces(img, ppath=pp, opath=op, rpath=rp)
except Exception:
print("Image is discarded due to exception!")
return 4
if len(
landmarks
) == 0: # If the landmarks cannot be detected, the img will be discarded
print("Image is discarded due to non-detected landmarks!")
return 4
if filter:
return True
else:
facial5points = [[landmarks[0][j], landmarks[0][j + 5]]
for j in range(5)]
warped_face = warp_and_crop_face(np.array(img),
facial5points,
reference,
crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
return img_warped
source_root = args.source_root # specify your source dir
dest_root = args.dest_root # specify your destination dir
crop_size = args.crop_size # specify size of aligned faces, align and crop with padding
scale = crop_size / 112.
reference = get_reference_facial_points(default_square = True) * scale
cwd = os.getcwd() # delete '.DS_Store' existed in the source_root
os.chdir(source_root)
os.system("find . -name '*.DS_Store' -type f -delete")
os.chdir(cwd)
if not os.path.isdir(dest_root):
os.mkdir(dest_root)
for image_name in tqdm(glob("{}/*".format(source_root))):
print("Processing\t{}".format(image_name))
img = Image.open(image_name)
try: # Handle exception
_, landmarks = detect_faces(img)
except Exception:
print("{} is discarded due to exception!".format(image_name))
continue
if len(landmarks) == 0: # If the landmarks cannot be detected, the img will be discarded
print("{} is discarded due to non-detected landmarks!".format(image_name))
continue
facial5points = [[landmarks[0][j], landmarks[0][j + 5]] for j in range(5)]
warped_face = warp_and_crop_face(np.array(img), facial5points, reference, crop_size=(crop_size, crop_size))
img_warped = Image.fromarray(warped_face)
img_warped.save(os.path.join(dest_root, image_name.split('.')[0] + '.jpg'))
arcface_r50_asian.to(device)
crop_size = 112
scale = crop_size / 112.
reference = get_reference_facial_points(default_square=True) * scale
# face detector
face_detector = FaceDetector()
image_1 = cv2.imread('./4/id.png')
image_2 = cv2.imread('./4/selfie.png')
bboxes_1 = face_detector.detect(image_1)
bboxes_2 = face_detector.detect(image_2)
facial5points_1 = [[bboxes_1[0][2*i - 1], bboxes_1[0][2*i]] for i in range(3, 8)]
warped_face_1 = warp_and_crop_face(cv2.cvtColor(image_1, cv2.COLOR_BGR2RGB), facial5points_1, reference, crop_size=(crop_size, crop_size))
facial5points_2 = [[bboxes_2[0][2*i - 1], bboxes_2[0][2*i]] for i in range(3, 8)]
warped_face_2 = warp_and_crop_face(cv2.cvtColor(image_2, cv2.COLOR_BGR2RGB), facial5points_2, reference, crop_size=(crop_size, crop_size))
"""
input_1 = preprocess(warped_face_1)
input_2 = preprocess(warped_face_2)
"""
embedding_1 = arcface_r50_asian(warped_face_1)
embedding_2 = arcface_r50_asian(warped_face_2)
cosin = cosine_similarity(embedding_1, embedding_2)
print("cosin:", cosin)
distance = euclidean_distances(embedding_1, embedding_2)
print('distance:', distance)
for image_name in os.listdir(os.path.join(source_root, subfolder)):
print("Processing\t{}".format(
os.path.join(source_root, subfolder, image_name)))
img = cv2.imread(os.path.join(source_root, subfolder, image_name))
try:
bboxes = face_detector.detect(img)
except Exception:
print("{} is discarded due to exception !".format(
os.path.join(source_root, subfolder, image_name)))
continue
if len(bboxes) == 0:
print("{} is discarded due to exception!".format(
os.path.join(source_root, subfolder, image_name)))
continue
count += 1
facial5points = [[bboxes[0][2 * i - 1], bboxes[0][2 * i]]
for i in range(3, 8)]
warped_face = warp_and_crop_face(cv2.cvtColor(
img, cv2.COLOR_BGR2RGB),
facial5points,
reference,
crop_size=(crop_size, crop_size))
cv2.imwrite(
os.path.join(dest_root, subfolder,
str(count) + ".jpg"),
cv2.cvtColor(warped_face, cv2.COLOR_RGB2BGR))
if count == 10:
break
