def opencv_cnn(image, **kwargs):
if is_grayscale(image):
raise SettingNotSupportedError("Grayscale not supported for " +
opencv_cnn.__name__)
model_file = "models/face_detection/opencv_face_detector_uint8.pb"
config_file = "models/face_detection/opencv_face_detector.pbtxt"
detector = cv2.dnn.readNetFromTensorflow(model_file, config_file)
scale_x = image.shape[1] / 300
scale_y = image.shape[0] / 300
def convert_rect(rect):
left, bottom, right, top = rect[3:7] * 300
return [
int(left * scale_x),
int(right * scale_x),
int(top * scale_y),
int(bottom * scale_y)
]
def run():
# part of benchmark because this is extra work that we would not do when not using this model
image_resized = cv2.resize(image, (300, 300))
image_scaled = cv2.dnn.blobFromImage(image_resized, 1.0, (300, 300),
(104.0, 177.0, 123.0))
detector.setInput(image_scaled)
result = list(detector.forward()[0, 0])
if (len(result)) == 0:
raise ExperimentFailure("No faces found by " + dlib_hog.__name__)
return [convert_rect(r) for r in result]
return run
def dlib_5point(image, **kwargs):
if is_grayscale(image):
raise SettingNotSupportedError("Grayscale not supported for " +
dlib_5point.__name__)
landmark_model = dlib.shape_predictor(
"models/face_cropping/shape_predictor_5_face_landmarks.dat")
face = get_first_face_location(image)
def run():
shape = landmark_model(image, face)
return dlib.get_face_chip(image, shape)
return run
def dlib_resnet(image, **kwargs):
if is_grayscale(image):
raise SettingNotSupportedError("Grayscale not supported for " +
dlib_resnet.__name__)
landmark_model = dlib.shape_predictor(
"models/face_cropping/shape_predictor_5_face_landmarks.dat")
feature_model = dlib.face_recognition_model_v1(
"models/feature_extraction/dlib_face_recognition_resnet_model_v1.dat")
face = get_first_face_location(image)
shape = landmark_model(image, face)
crop = dlib.get_face_chip(image, shape)
def run():
return feature_model.compute_face_descriptor(crop)
return run
def opencv_haar(image, **kwargs):
detector = cv2.CascadeClassifier(
'models/face_detection/haarcascade_frontalface_default.xml')
# not part of benchmark because would supply image in correct color order in main code
if not is_grayscale(image):
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
def convert_rect(rect):
x, y, w, h = list(rect)
return [x, x + w, y + h, y]
def run():
result = list(
[convert_rect(r) for r in detector.detectMultiScale(image)])
if (len(result)) == 0:
raise ExperimentFailure("No faces found by " + dlib_hog.__name__)
return result
return run
def openface(image, **kwargs):
if is_grayscale(image):
raise SettingNotSupportedError("Grayscale not supported for " +
dlib_resnet.__name__)
landmark_model = dlib.shape_predictor(
"models/face_cropping/shape_predictor_5_face_landmarks.dat")
feature_model = cv2.dnn.readNetFromTorch(
"models/feature_extraction/nn4.small2.v1.t7")
face = get_first_face_location(image)
shape = landmark_model(image, face)
crop = dlib.get_face_chip(image, shape)
def run():
crop_blob = cv2.dnn.blobFromImage(crop,
1.0 / 255, (96, 96), (0, 0, 0),
swapRB=True,
crop=False)
feature_model.setInput(crop_blob)
return feature_model.forward()[0]
return run