def test_align():
image = cv2.imread("timg4.jpg")
bounds, lmarks = gen_face(MTCNN_DETECT, image)
crops = MTCNN_DETECT.extract_image_chips(image, lmarks, padding=0.4)
if len(crops) == 0:
raise Exception("no crops~~ %s" % image_path)
bounds, lmarks = gen_face(MTCNN_DETECT, crops[0])
org_box, first_lmarks = bounds[0], lmarks[0]
trible_box = gen_boundbox(org_box, first_lmarks)
pitch, yaw, roll = get_rotation_angle(image, first_lmarks)
print(pitch, yaw, roll)
cv2.imwrite("test.jpg", crops[0])
def crop_and_trans_images(self, detector, series):
# imdb 数据存在多张人脸，所以对于多人脸的数据直接清除掉
image_path = os.path.join(self.data_dir, series.full_path[0])
try:
print(image_path)
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
if not np.isnan(series.second_face_score):
raise Exception("more than one face~---%s~-%s- %s" %
(series.name, series.age, image_path))
bounds, lmarks = gen_face(detector, image, image_path)
crops = detector.extract_image_chips(
image, lmarks,
padding=0.4) # aligned face with padding 0.4 in papper
if len(crops) == 0:
raise Exception("no crops~~ %s---%s" %
(image_path, series.age))
if len(crops) > 1:
raise Exception("more than one face~---%s~-- %s" %
(series.name, image_path))
bounds, lmarks = gen_face(detector, crops[0],
image_path) # recaculate landmar
org_box, first_lmarks = bounds[0], lmarks[0]
trible_box = gen_boundbox(org_box, first_lmarks)
pitch, yaw, roll = get_rotation_angle(
crops[0], first_lmarks) # gen face rotation for filtering
image = crops[0] # select the first align face and replace
except Exception as ee:
logging.info("exception as ee: %s" % ee)
trible_box = np.array([])
org_box, first_lmarks = np.array([]), np.array([])
pitch, yaw, roll = np.nan, np.nan, np.nan
age = np.nan
gender = np.nan
status, buf = cv2.imencode(".jpg", image)
series["image"] = buf.tostring()
series["org_box"] = org_box.dumps() # xmin, ymin, xmax, ymax
series["landmarks"] = first_lmarks.dumps() # y1..y5, x1..x5
series["trible_box"] = trible_box.dumps()
series["yaw"] = yaw
series["pitch"] = pitch
series["roll"] = roll
return series
def loadData_preprocessData_and_makeDataFrame():
meta_dataframe = pd.read_csv(
dataset_base_path.joinpath(dataset_name + '_meta.csv'))
properties_list = [] # init lists of all properties gonna be saved
# loop through meta.csv for all images
for index, series in meta_dataframe.iterrows():
image_path = series.full_path # get image path
try:
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
# image = cv2.copyMakeBorder(image, self.extra_padding, self.extra_padding, self.extra_padding, self.extra_padding, cv2.BORDER_CONSTANT)
face_count, _, lmarks_list = detect_faces_and_landmarks(
image) # Detect face & landmarks
if face_count != 1:
raise Exception("more than 1 or no face found in image ",
image_path)
# found exactly 1 face, so now process it
#extract_image_chips will crop faces from image according to size & padding and align them in upright position and return list of them
cropped_faces = dlib.get_face_chips(
image, lmarks_list, padding=extra_padding
) # aligned face with padding 0.4 in papper
image = cropped_faces[0] # must be only 1 face, so getting it.
# detect face landmarks again from cropped & align face. (as positions of lmarks are changed in cropped image)
_, face_rect_box, lmarks_list = detect_faces_and_landmarks(
image) # Detect face from cropped image
first_lmarks = lmarks_list[
0] # getting first face's rectangle box and landmarks
trible_box = gen_boundbox(
face_rect_box, first_lmarks
) # get 3 face boxes for nput into network, as reauired in paper
if (trible_box < 0).any():
print(index, 'Some part of face is out of image ',
series.full_path)
raise Exception("more than 1 or no face found in image ",
image_path)
face_pitch, face_yaw, face_roll = get_rotation_angle(
image, first_lmarks) # gen face rotation for filtering
except Exception as ee:
# print('index ',index,': exption ',ee)
properties_list.append([
np.nan, np.nan, np.nan, np.nan, np.nan, np.nan, np.nan, np.nan,
np.nan, np.nan
]) # add null dummy values to current row & skill this iteration
continue
# everything processed succefuly, now serialize values and save them
status, buf = cv2.imencode(".jpg", image)
image_buffer = buf.tostring()
#dumping with `pickle` much faster than `json` (np.dumps is pickling)
face_rect_box_serialized = face_rect_box.dumps(
) # [xmin, ymin, xmax, ymax] : Returns the pickle(encoding to binary format (better than json)) of the array as a string. pickle.loads or numpy.loads will convert the string back to an array
trible_boxes_serialized = trible_box.dumps(
) # 3 boxes of face as required in paper
landmarks_list = np.array(
[[point.x, point.y] for point in first_lmarks.parts()]
) # Same converting landmarks (face_detection_object) to array so can be converted to json
face_landmarks_serialized = landmarks_list.dumps(
) #json.dumps(landmarks_list,indent = 2) # y1..y5, x1..x5
# adding everything to list
properties_list.append([
image_path, series.age, series.gender, image_buffer,
face_rect_box_serialized, trible_boxes_serialized, face_yaw,
face_pitch, face_roll, face_landmarks_serialized
])
if index % 500 == 0:
print(index, 'images preprocessed')
processed_dataset_df = pd.DataFrame(properties_list,
columns=[
'image_path', 'age', 'gender',
'image', 'org_box', 'trible_box',
'yaw', 'pitch', 'roll', 'landmarks'
])
# some filtering on df
processed_dataset_df = processed_dataset_df.dropna()
processed_dataset_df = processed_dataset_df[
(processed_dataset_df.age >= 0) & (processed_dataset_df.age <= 100)]
# processed_dataset_df.to_csv('/content/Dataset.csv',index=False)
#Dataset_DF = processed_dataset_df # putting it into global Dataset_DF variable
return processed_dataset_df # returning now (just in case need to return), maybe later remove...