def detect_face_boxes(self, img: np.ndarray) -> List[Box]:
img_300x300: np.ndarray = cv.resize(src=img,
dsize=DETECTION_NETWORK_INPUT_SIZE)
img_caffe_blob: np.ndarray = cv.dnn.blobFromImage(
image=img_300x300,
scalefactor=1.0,
size=DETECTION_NETWORK_INPUT_SIZE,
mean=BLOB_MEAN_SUBTRACTION.to_bgr(),
)
self.face_detection_network.setInput(img_caffe_blob)
inferences: np.ndarray = self.face_detection_network.forward()
# Parsing inferences informed by:
# https://answers.opencv.org/question/208419/can-someone-explain-the-output-of-forward-in-dnn-module/
confidences: Sequence[float] = inferences[0, 0, :, 2]
y_scale, x_scale, _ = img.shape
# 1. Extract box coordinates (all between 0 and 1)
# (top_left_x, top_left_y, bottom_right_x, bottom_right_y)
# 2. Scale them up to original image size
# 3. Round values and convert to int
# 4. Collect to Box
boxes: List[Box] = [
Box.from_values(*values)
for values in ((inferences[0, 0, :, 3:7] *
np.array([x_scale, y_scale, x_scale, y_scale])
).round().astype(int))
]
return [
box for confidence, box in zip(confidences, boxes)
if confidence >= self.detection_threshold
]
def gender_faces(img: np.ndarray, faces: List[Face]) -> List[GenderedFace]:
"""
Classify the genders of the faces in an image
:param img: The image
:param faces: The bounding boxes of the faces in img
:return: List of inferred genders and the confidence value parallel with face_boxes
"""
ret = []
pad_width = max(img.shape) + 100 # TODO: Make this smarter
img = cv2.copyMakeBorder(img, pad_width, pad_width, pad_width, pad_width,
cv2.BORDER_REFLECT_101)
for face in faces:
# Extract each face, and resize to (H,L)
center_x, center_y = face.box.center
h = face.box.bottom_right_y - face.box.top_left_y
w = face.box.bottom_right_x - face.box.top_left_x
side_len = max(w, h)
radius = side_len / 2
box = Box.from_values(
center_x - radius + pad_width,
center_y - radius + pad_width,
center_x + radius + pad_width,
center_y + radius + pad_width,
)
face_crop = img[box.top_left_y:box.bottom_right_y,
box.top_left_x:box.bottom_right_x, :]
face_crop = cv2.resize(face_crop, (H, L))
# Extract features. If features are not found, use defaults.
feat = face_utils.shape_to_np(face.shape)
landmarks = np.zeros(8)
x = box.top_left_x
y = box.top_left_y
landmarks[0] = feat[3][0] - feat[2][0] - x
landmarks[1] = feat[1][0] - feat[0][0] - x
landmarks[2] = feat[4][0] - x
landmarks[3] = feat[4][0] - x
landmarks[4] = feat[3][1] - feat[2][1] - y
landmarks[5] = feat[1][1] - feat[0][1] - y
landmarks[6] = feat[4][1] - y - 5
landmarks[7] = feat[4][1] - y + 5
# Classify
X = np.concatenate(
(face_crop.reshape(L * H * 3), landmarks)).reshape(1, -1)
y_prob = GENDER_CLASSIFIER.predict_proba(X)[0]
y_pred = GENDER_CLASSIFIER.predict(X).astype(int)[0]
# Create GenderedFace object
g_face = GenderedFace.from_face(face=face,
gender=Gender(y_pred),
gender_confidence=y_prob[y_pred])
ret.append(g_face)
return ret